I heard about the FOP programme at Regeneron the way you hear about most things in a big company, occasionally and in passing. Aris Economides gave a talk on it a couple of times. I missed both, then heard from my colleagues that it was inspiring, and that FOMO drove a curiosity in me about what the story actually was. Then I procrastinated.

When the trial results were press released there was excitement in the building. A rare disease programme hitting a milestone like that is not an ordinary day. My curiosity grew, and I decided I should write about it. Because, you know, writing is how I discipline myself into learning something properly. Then I procrastinated again. I told myself I would wait for the trial to be published so I could time the post to it, and then I forgot about it entirely.

I left the company. I moved to India. And a few weeks ago I learnt that the FDA approval came through, before the trial results were ever published.

That was the trigger.

So I have spent the past weeks reading, and I should clear something first. I was at Regeneron while this programme was taking final shape, but I was nowhere close to it, and I have no insider information whatsoever. Nothing here comes from a corridor conversation, and my sources are Aris's own review article and a long list of papers from the literature. The story naturally turned out to bear an outsider's perspective.

I wish I had done this homework while I was still there, so I could have stopped Aris at any of the times I crossed paths with him and asked him about it directly. But I think it works better this way. What follows is what the published record alone will give you.

And the reason I could not let it go is that this story is not really about human genetics, which is my narrow field of vision, even though that is of course at the core of it. It is about what it takes to develop a drug for a puzzling, debilitating illness. That is a perspective I think anyone who imagines themselves doing that one day ought to have.

The news

On 19 August 2026 the FDA approved garetosmab, developed by Regeneron for a rare disease called fibrodysplasia ossificans progressiva (FOP). Garetosmab, a monoclonal antibody against a protein called activin A, is the second drug ever approved1 for this disease, but it is the next best thing to a cure these patients can have, shown by the outcome of the phase 3 trial.

The disease

FOP is an extremely rare disease with estimates of around 900 diagnosed across the world. It's a bone-forming disease. The telltale sign appears at birth: malformed great toes, characteristic of the condition. That remains the only sign for years.

The bone comes later, usually in childhood, preceded by an injury, a fall, an injection, sometimes nothing one can identify. A lump appears and it hardens. Don't mistake it for scar tissue with calcium in it. It is real bone, laid down through a cartilage stage, the same route by which a growing child builds a femur. It happens in muscle or tendon or ligament, in a place where a skeleton is not supposed to be.

It happens in episodes. A flare, then months or years of nothing, then another bone somewhere else. Between them the disease is silent.

The formed bone never goes away, and every episode adds to what is already there. Joints fuse. The spine locks. Most are in a wheelchair by thirty, and half do not live past their mid fifties. Death is often painful. It comes from an immobilised chest wall, neck and jaw that slowly halt swallowing and finally breathing.

The obvious thing to do with a lump of bone in the wrong place is to cut it out. In FOP you must not. Resection puts back more bone than it removed. Surgery is contraindicated. So is a biopsy or even an intramuscular injection. Every ordinary instrument medicine has for looking at a lesion or for getting a drug into a muscle is a way of making the disease worse.

That shuts down most of the options for intervening. Nothing helps once the bone is formed. Anything that works (including garetosmab) has to work before the bone starts.

The company

First things first. A hearty congratulations to team Regeneron, especially Aris Economides, vice president who runs its skeletal diseases group. Aris joined the company straight out of his PhD, and has been there ever since. He championed this programme, and Regeneron moved the molecule through phase 3 to approval for a disease with a mere 900 diagnosed patients in the world.

On first impression this could look easy. An antibody against a protein, made by a company that is a pioneer at making antibodies. A rare disease, a protein you can block, and a manufacturer already equipped to block it. A company without deep pockets has to bet everything on it, and the market is too small to build a business around. A large one can absorb it, and there is often a priority review voucher at the end of it, which can be worth more than the drug will ever earn.

That is a reasonable thing to assume, and it is not what happened here.

You see, the antibody was never the hard part. It's the target. Uncovering the right target is the story. And the reason a company with no particular reason to be in this disease went deep into it is serendipity and curiosity more than a strategic move. Regeneron happened to be holding an antibody against activin A, or was planning to make one, for something entirely unrelated to bone. It then appeared that there might be another use for it. That use was not straightforward, and somebody decided it was a challenge worth undertaking.

The timeline

One of the things that I pay attention to when reading about a new medicine is the timeline. The commonly quoted ones measure the time from when a gene was mapped to a disease to the date of FDA approval. That measurement is what answers the most asked question, how long does it take to develop a drug. By that standard it is 20 years for garetosmab. The gene was mapped in 2006 and the drug was approved in 2026.

But that shortsighted timeline misses some important history, and it understates how long it really took for a cure to arrive for the patients. As you unroll the timeline, you'll see that it keeps going back. 48 years, if you start from the assembly of the protein family that activin A belongs to, built out of nothing but sequence homology, by people in four unrelated fields. 61 years, if you start the clock from the experiment that showed bone can be induced in places where no bone belongs, done in a rabbit in 1965 by someone who had never heard of this disease. 286 years, if you start from the first clear description of the disease itself, in a letter to the Royal Society published in 1740, describing a boy whose ribs had joined across his back.

That is what solving a disease actually costs. And almost none of that was spent by people working on the disease.

What follows below is a glimpse of that longer timeline, not a full history of it. That's too big to fit in here.

It is worth knowing how the story is structured. The story stem winds around the disease, from the first description to the drug. But then there are branches that join at points where the disease goes quiet and something in the background takes over. That is not a digression. That is what FOP was waiting for. And none of it looked like a line at the time. It only becomes one backwards.

Before travelling back, let me give you a snapshot of where we are heading. It's not a spoiler. It's not what the answer turned out to be. It's rather how anyone came to know it.

FOP is caused by a single nucleotide change in the gene ACVR1. It encodes a receptor protein, named activin A receptor type I, which is part of a superfamily called TGF-beta. There are 12 receptors and about 33 ligands in this family, which branches into two pathways, each orchestrating different biology. ACVR1 sits in the branch that builds bone and it binds to a bone morphogenetic protein (BMP). Activin A is not a BMP.

The first record

On 14 April 1736 a 14 year old boy walked into St Bartholomew's Hospital in London and asked what could be done for him. John Freke, a surgeon there, wrote down what he saw.

There came a boy of healthy look and 14 years of age, to ask of us at The Hospital, what should be done to cure him of many large swellings on his back which began about three years since, and have continued to grow as large on many parts as a penny-loaf, particularly on the left side. They arise from all the vertebrae of the neck, and reach down to the os sacrum. They likewise arise from every rib of his body, and joining together in all parts of his back, as the ramifications of coral do, they make, as it were, a fixed bony pair of bodice.

Two things to pay attention to in Freke's vivid description. The timing: a healthy looking boy who noticed swellings growing on his back for three years. He was an ordinary child until he was eleven. The disease waited, and then it started. The coral: the swellings are not separate lumps on the boy's back. They arose from every rib and joined across all parts of it, becoming one continuous piece, closing around him like bodice. There was never a discrete thing to cut away. Nothing clinical has been added to that record in 286 years.

For most of those years the disease had the wrong name, myositis ossificans progressiva. The problematic part is 'myositis', which means inflammation of muscle. The disease was believed to be a disease of the muscle for nearly two and a half centuries.

If you think about it, it was a reasonable guess. The lumps come up in muscle, often after an injury, and they are hot and swollen while they appear, and there was no way to examine the swelling beyond what you could see at the patient's bedside.

In 1972 Victor McKusick struck out 'myositis'. He said the bone does not come from muscle. It comes from the connective tissue running through and around it, and the muscle is an innocent bystander. He replaced myositis with fibrodysplasia, and it has been fibrodysplasia ossificans progressiva ever since.

McKusick had the tissue right in 1972, and then nothing happened for thirty four years.

It was not that no one wanted to try anything. Asking which gene causes a disease was not a question that anyone could ask by then. Finding a human disease gene from its position on a chromosome, with no idea in advance what that gene might be, did not exist as a technique until the mid-1980s.

When the gene mapping technique did arrive, there was a second problem waiting already. Linkage needs families, ideally several generations, affected and unaffected in each, so you can watch a stretch of chromosome travel down the family alongside the disease. FOP immobilises people in childhood and adolescence, and people with FOP rarely have children. So there were no multigenerational families to map the genetics.

So the disease waited. But the vocabulary it was waiting for was already being built, in laboratories that had never heard of FOP. In 1965, seven years before McKusick fixed the disease name, a surgeon working on why bone grafts heal had already reproduced the core of FOP's disease process in a rabbit and published it.

The family

In 1965, Marshall Urist, an orthopaedic surgeon at UCLA, was trying to answer an ordinary question that had puzzled bone surgeons for eighty years. Bone grafts work. Why? The assumption since 1881 was that the graft's own living cells make the new bone, so a graft is a delivery vehicle for bone forming cells.

Urist took bone, killed anything living in it by dissolving out its mineral with acid, and implanted the residue2 into a rabbit's muscle. Bone formed. In the muscle, where no bone belongs, through a cartilage stage, the same route by which a growing child builds a femur. No donor cells could make it as there weren't any. So the new bone had to be coming from host cells, converted into bone-makers by something in the debris. He called it autoinduction, and six years later he named that something bone morphogenetic protein.

Note, the new idea there is induction, not ossification, which is already a recognised concept. What was new is a signal that bestows a new power, bone-making capability, on cells that could never do it before.

Now reread Urist's experiment as a clinical description. Cell-free material placed in skeletal muscle, and endochondral bone comes out of it. That is FOP. In 1965, forty one years before anyone knew which gene was involved, the core disease process was remarkably mimicked in a rabbit by a man who had never heard of FOP. The disease was demonstrated long before it could be studied.

And that is the pathway the FOP field would be fixated on eventually. Which is what made the next forty years of wrong answers possible.

While the bone people were doing that, three strangers were building something none of them meant to build.

It starts in 1978 with two cancer virologists. De Larco and Todaro were working on cells transformed by a mouse sarcoma virus, and they found those cells were secreting something that made ordinary fibroblasts behave like cancer cells, growing in soft agar where normal cells will not. They called it sarcoma growth factor. It was later split into two, and one half of it became transforming growth factor beta. The naming is after an artefact in a dish.

Its gene sequence was published in August 1985. Four months later, in December, a reproductive endocrinologist named Mason published the sequence of inhibin, the fertility hormone, and noticed it looked like the cancer factor. He reported these are members of one gene family. That sentence is where the superfamily begins. In 1987, a fly geneticist working out how a fly embryo decides which side is its back found that his gene was another relative. And in December 1988 the bone morphogenetic proteins were cloned and filed in as members of the same family, twenty three years after Urist's rabbit experiment.

Look at what those four people had in common. Nothing. Not a tissue, not an organ, not a function, not a disease, not even an assay. A cancer experiment, a fertility hormone, a fly's embryo, and a protein from digested bone. The only thing that connected them was that their protein sequences resembled each other. The entire family was assembled out of sequence and nothing else.

Which is also why the family is named after the wrong thing. Whichever member gets sequenced first becomes the one everything else is compared to, and, in this particular case, it was a matter of four months. Transforming growth factor beta was published on 22 August 1985, and inhibin on 19 December 1985. Reverse that, and the family would be called the inhibin family. Urist's bone protein was filed in 1988 under a name that came from cells growing in soft agar.

And in 1988 that is all there is. A family, and no map of it. Nobody knew which member spoke to which, or that the family divides into two halves that do totally different things. The receptors were yet to arrive. When they did, each one would be named after whichever ligand it seemed to bind, and nobody had any way of telling a right name from a wrong one.

Activin arrives in this story the way most things in it arrive, sideways.

In 1932 an American physician called McCullagh injected a water extract of testis into castrated rats and observed what happened to the pituitary. Castration normally releases the brake that restrains the gonad-stimulating hormones secreted in the pituitary, causing the pituitary cells to swell and go vacuolated. Pathologists of that era called them castration cells. McCullagh noticed that the injection prevented the castration response in the rat's pituitary. And it did nothing for the prostate. He inferred that whatever that extract contained was not a male hormone but something else that inhibited the pituitary response. He called it inhibin.

Then fifty years of nothing. Inhibin was found to be a real substance in bioassays but nobody knew what it was exactly. The protein was purified finally in 1985, out of pig follicular fluid. It turned out that ovaries, more than testes, are the more abundant source and you get litres of it from a slaughterhouse. And in June 1986 two laboratories purifying the protein encountered a fraction of the extract doing the exact opposite of what they expected: stimulating the pituitary rather than suppressing it. It turned out it was the same protein but different subunit configuration. Inhibin is an alpha bound to beta subunit. When beta binds to beta, you get the inverse action. And that's how activin was born, out of an accident.

By 1990 activin had left reproductive biology altogether. The factor that tells a frog embryo which of its cells become muscle and bone turned out to be activin, and graded doses of it gave graded changes in the cells' fate, which is how a morphogen behaves. So the people who went looking for activin's receptor were developmental biologists. Nobody in that hunt was thinking about bone, and nobody had ever seen a receptor for any member of this family.

They found one in 1991, by the oldest method that existed. Mathews and Vale, from the same laboratory that discovered activin, labelled activin A with radioactive iodine and used it as bait to fish out whatever it gripped on the cell surface. What was caught was ACVR2A, the very first receptor of the entire superfamily, and it came with a surprise. It was a serine threonine kinase, when every growth factor receptor the field knew before was a tyrosine kinase. A new class of receptor was thus born, pulled out using a hormone from an ovary.

And that is the last time anyone found a receptor in this family by its ligand. A second one came a year later by looking for sequences that resembled the first. Then in 1993 the same trick at scale: take the conserved stretch of that kinase, go fishing with the sequence bait instead of the hormone. Four receptors caught in the net at once, and the series ran to seven within three years. They were caught blind. Nobody knew what a single one of them bound. They were named for the thing they resembled, and the naming is honest if you read it literally, activin receptor-like kinases, ALK1 through ALK7. The resemblance that had assembled the ligands was now assembling the receptors, and it was doing it much faster than anyone could figure out what they were for. To highlight one of them, ALK5 is the TGF-beta receptor.

Then the orphan receptors were tested against ligands, and two of them held on to activin. One was ALK4. The other was ALK2, which is ACVR1, the FOP gene. Both took activin into their names on that evidence, and the evidence was real. ALK4 was later found to pass the signal on into the cell. But ALK2 did nothing. It just bound activin. But these were realised only later and not known in 1993, because binding was the only thing anyone could measure then. The machinery inside the cell that carries the signal onward was not found until 1995 and 1996. Until then you could only tell if a receptor binds to something. You could not tell if that binding did anything to the cell.

So ACVR1 was discovered by sequence resemblance and named by binding, and at that moment there was nothing in existence that could have shown the naming to be wrong. The signalling map came afterwards. When it did, the family split into two halves that answer to different sets of ligands and handle different machinery inside the cell. ACVR1 sat on the bone half while activin sat on the other. A name given in 1993 (ACVR1) was checked against the signalling map drawn in 1996, and it has been treated as a misnomer ever since.

The goose chase

While all of that was happening in other people's laboratories, the FOP field had not gone quiet. Through the 1980s and the 1990s, Fred Kaplan and Michael Zasloff at the University of Pennsylvania were doing work that looks like administration rather than science. They were finding FOP patients. A disease this rare drops one at a time, scattered across countries. Usually they get misdiagnosed at first, so they took referrals from anywhere, wrote to families and slowly assembled an international register of people with FOP.

The first real finding came in 1996, appearing in the New England Journal of Medicine. Shafritz, Shore and Kaplan took cells from FOP patients and measured the bone morphogenetic proteins, BMP-1 through BMP-7 one after the other, looking for a protein that the patients' cells made too much of. One of them turned up. BMP-4 was found to be expressed more in patients' cells than controls'. The suspicion was older than the paper, and it says so in the introduction, that too much BMP had already been proposed as the cause of the disease. What was new after the NEJM paper is that there was now a number under it, BMP-4.

Note which half of the system they searched, the ligands. That was the only half you could search. By the mid 1990s the BMPs were a numbered list, and they were cloned and catalogued, so you could take them one at a time and measure. The receptors though were not a list yet. They had been fished out blind just three years back, half of them still had no confirmed ligand, and there was nothing to read them out with, since the machinery inside the cell they signalled was found only in 1995 and 1996. You could ask what the patients' cells were making too much of. But you could not yet ask what the receptors on their surface were talking to inside the cell.

What followed was ten years of a wild goose chase. Every research effort built on the finding from the NEJM paper. If the patients' cells make too much BMP-4, the fault should be in the gene for BMP-4. They sequenced it. Nothing. If the protein is not built wrong, then perhaps it is switched on wrong, so in 1998 they worked out the structure and the transcriptional control of the human BMP4 gene. Nothing there either. If the ligand is not made wrong and not controlled wrong, perhaps its brake is broken, so they went to a protein called noggin that holds BMPs down, BMP-4 among them, and that is a story of its own and comes next. And if not the ligand, its gene, its control or its brake, then the receptor that binds it, BMPR1A, BMP-4's own receptor, which in 2005 turned out to be trafficked abnormally in FOP cells and that carried no mutation either. Ten years of search, and the true FOP gene was nowhere in the vicinity of that search space.

To be fair, none of that is foolish. Each research effort was a rung laid above the one below it in the BMP-4 ladder that was brought to the scene by the NEJM paper. The issue was the BMP-4 finding at the bottom of the ladder was not a false one. By 2003 they did find that BMP-4 was transcribed five to seven times faster in patient cells than in control cells, though they couldn't say why. As far as anyone could tell, it was a real finding. A wrong measurement gets dropped in a year or two, and it will not come back. A right measurement pointing at the wrong part of the pathway can run futile for a decade. And that's what happened.

Now look at what was on the list they measured: all BMPs one to seven and nothing else. The protein that actually drives this disease was outside the search before the search began. It doesn't matter anyway, as a broader list wouldn't have helped either. Activin A is not raised in FOP patients either, not during a flare and not between flares. Nothing in these patients is being made in excess. The assumption underneath the entire ladder was that a disease of too much bone has to come from too much signal, and that logically sound assumption turned out to be wrong in the FOP puzzle.

The BMP4 ladder had ended on a receptor, BMPR1A, which led to nothing. BMPR1A belongs to the BMP type I receptors, a group with only three members. The FOP gene is one of them. The goose chase after all took the field to the right shelf, but just made them pick the wrong neighbour. The right one had been there since 1994. ACVR1 had been fished out blind with the other orphans, had bound activin, had been named for it, and then it entered the BMP family in 1994, when the BMPs were tested against the same seven receptors, and ACVR1 was found to bind BMP-7. The same sentence that assigns ACVR1 to BMP-7 says explicitly that it doesn't bind BMP-4. So the gene that causes FOP was already in its home, the BMP receptor family, when the field was hunting for the gene in the BMP space.

Noggin is not a member of the BMP family. It is a secreted protein that shuts down BMPs from outside the cell. Knowing how it does that is important as the idea comes back at the end. A BMP works by clutching two receptors at once, a type I and a type II. Noggin hugs BMP around covering exactly the two spots that connect the BMP to the receptors. Purely mechanical. Nothing is destroyed or switched off. The BMP is still there in the tissue but incapacitated, unable to talk to the receptors.

So, noggin earned its entry into the FOP story through logic. If a disease of too much bone doesn't come from making too much BMP, the same excess could come from losing the thing that restrains it. That idea needs no fault in any BMP gene at all, which was convenient, because the FOP field had sequenced BMP-4 and found nothing. Noggin was the brake, and a broken brake would look exactly like a stuck accelerator.

The claim arrived in 1999. A French group reported that an FOP patient carried a 42 base pair deletion in the noggin gene, and the year after, the same group published a linkage map to the chromosome where the noggin gene sits. Two results, arriving separately, agreeing with each other and with the hypothesis everyone already held. On paper that is what a solved disease looks like.

Penn tested the noggin claim across thirty one families, starting with a linkage analysis, and found no evidence. They even sequenced the gene itself, in four families and eighteen patients, and found no mutations. They published the null result in 2000, one year after the claim. The question of noggin as the FOP gene should have closed, except it didn't.

In 2001 three further noggin mutations were reported, in three more families. In 2002, Penn published a paper whose title says the reported noggin mutations are PCR errors. Yet new noggin claims were continuing to appear in 2005. In 2008, two years after the real FOP gene had already been found and published, a paper appeared claiming mutations in noggin and in ACVR1 together, and drew two separate rebuttals in the same issue of the journal. One of those rebuttals carries about twenty names, and one of them is Aris Economides. That is how he enters this story, and it is not because he worked on FOP. His first paper, in 1993 in Science, was on noggin, and he had spent the fifteen years since on the proteins that hold BMPs down, and nothing but BMPs. He was there as an expert on the protein under trial, the protein that was never the answer to the FOP puzzle.

The final nail in the noggin coffin was none of the rebuttals but the very first patient in whom the noggin mutation was found, the case that ignited a second goose chase alongside the BMP-4 one. When the FOP gene and the causative mutation were eventually found (which is the next part of the story), the noggin patient zero turned out after all to be a case of ordinary FOP carrying the ACVR1 mutation. The noggin mutation was an innocent bystander. Nobody had to argue it down.

By 2000 the first FOP genetic maps were starting to come in. In January 2000, a Penn group took the four FOP families and drew a linkage map, which pointed to a locus on chromosome 4 with the LOD score that just surpassed the significance threshold. But more than that what increased the confidence on that locus was the authors' own closing line, that the interval contains at least one gene involved in the BMP signalling pathway. The map had landed where the hypothesis wanted it to land.

The material behind that score was four families. With that sample size, the strongest result the study could produce sits barely above what a false peak can reach. It also ran several hundred markers, so the odds of some random locus clearing the threshold were not small. The same year supplied the demonstration. The other genome scan of 2000, the linkage scan that produced the noggin locus, ran on seven families and produced a statistically significant LOD score. Two maps, in the same year, both above threshold and both wrong.

After the map came the gene hunt. Candidate genes inside the chromosome 4 interval were sequenced, and no mutations were found. Which genes those were was never recorded. Nor is it known whether the BMP pathway gene that made the interval attractive in the first place was among them.

Reading one gene across a handful of patients was ordinary work, and labs were doing exactly that outside the interval too, one BMP gene at a time on intuition, which is what the previous ten years of this section have been. One thing is certain though: ACVR1 was not among the genes anyone read. Had it been, the mutation was sitting there to be found, the same single base in every patient, and the search would have ended in an afternoon. What the era could not do was read the entire list. Twenty odd BMPs and growth factors, seven type I receptors, five type II, the antagonists, several hundred exons across dozens of patients, and one gene at a time was the unit of work. Sequencing everything was not a decision anyone refused to take. It was simply not possible. So the discovery had to wait for something to point the scientists at a locus on the long arm of chromosome 2, where ACVR1 was waiting to be sequenced.

The gene

The next linkage scan came from the same Penn group, and there was nothing remarkable about it. Five families, against four last time. No new method, no new markers. What was different is that every family in it had been handpicked to ensure a strict phenotype: malformed great toes at birth and bone forming in the characteristic pattern, in every affected family member. No ambiguous case was allowed into the study. They could afford to be that strict only because they had far more patients to start with, having assembled them one at a time over a period of twenty years.

What they found was another locus, this time on chromosome 2. It was no better than what they had seen before. In fact, it scored lower than either of the two wrong signals of 2000, and it did not reach the threshold for statistical significance. And like before, there was a BMP gene sitting inside the region, which is what made them consider this locus despite its weak statistical significance, with an intuition that perhaps they were close to the answer this time. They were.

The moment of truth arrived when they sequenced the candidate gene. Case after case carried the same mutation, same single base, in every patient they looked at. That one finding validated the locus, leaving no place for any doubts. Ten years of gene hunt finally ended. The gene behind FOP was finally found. It was ACVR1.

The mutation

The same base. Not just the same gene, the exact same base. Seven families carried it, and so did thirty two out of thirty two patients who had no family history at all. There are only two ways to interpret that. Either all of those people are distant relatives who inherited one ancient mistake, or that particular letter breaks on its own, over and over, in people who are not related.

They checked, and the answer couldn't be any clearer. The five families came from different ancestries, and the stretch of DNA carrying the mutation was different in every one of them. There is no common ancestor behind it. The nucleotide change was absent from 159 unaffected people and from the variant databases of the time. In six of the sporadic cases they sequenced both parents and found the mutation in none. So the mutation had appeared new in the child. The authors described it in the paper as one of the most specific codons in the human genome to be tied to a disease.

The reason is chemistry, and it arrived only three years later. The letter sits in a CpG, a C followed by a G, which is the most fragile pair in the genome. The C in that pair carries a methyl group, and a methylated C decays into a T3 faster than any other letter turns into anything else. So the mutation is not descent. It is a chemical accident happening again and again in people who have never met4.

The same chemistry also allows us to estimate the number of individuals carrying the mutation around the world. The mutation happens about once in every 10 million times a gamete is made, and the child can inherit it from either parent, so roughly one birth in five million. FOP shortens life, so at any moment fewer carriers are alive than were born. The calculation gives the prevalence as on the order of a thousand people alive on earth with FOP, close to the 900 diagnosed around the world.

The important thing to know is the mutation keeps coming. It is barely ever passed on to the next generation, so selection has no job here. It's purely the chemistry that replenishes the pool of patients continuously.

The hypothesis

To understand the mutation site and how the field thought through its mechanism, you need to know first how the receptor signalling works. The signalling starts from the ligand outside the cell, and it is a pair of halves, usually identical. It gathers four receptors around itself, two type II and two type I, and the assembly as a whole is what does the job. All four are transmembrane proteins of the same build: a small ligand-binding piece hanging outside the cell, one stretch crossing the membrane, and a kinase hanging inside. The type II kinase is on all the time, and its only job inside the cell is to switch on the type I receptor beside it. The type I receptor then switches on the downstream proteins, propagating the signal into the nucleus and turning on target genes. Which type I receptors assemble into the complex determines which of the two signalling pathways that define the two branches of the family switches on.

The switch that the type II kinase throws is a short stretch called the GS domain, on the type I receptor, just inside the cell, between the membrane and the kinase. It is the type I receptor's identity and is absent on type II. The switch shifts between two positions. Unphosphorylated, it folds back against the kinase and holds it shut. Phosphorylated, it becomes the dock where the signal lands. There is even a guard on the off position, a small protein that sits on the unphosphorylated box and keeps the type II kinase from firing it by accident. That is where the FOP mutation landed.

Nobody needed an experiment after that. The location alone gave the answer. The mutation is in the brake, and a mutation in a brake releases the brake. It was not even a guess. Eleven years earlier, point mutations in this same box in a different type I receptor had been shown to make it fire with no ligand present. By 1999 a permanently switched-on version of this very receptor, built the same way, was a standard reagent in labs. Nobody had to invent anything to believe the FOP mutation switched the receptor on.

The mutation landed in the one part of the receptor the field knew best, and that is what made everyone so confident about what it was doing. What the field knew was the job of the domain the mutation was sitting on. Not the mutation itself. Knowing what a brake does lets you form a hypothesis about the mutation in it. That is not a finding about the mutation. So the gap is not ignorance. They knew a great deal, and it was exactly enough to stop asking questions. There was a second blind spot underneath it. That part sits inside the cell, so all the attention went to what happens after a ligand has already docked on the outside. Nobody had a reason to ask what the receptor was answering to in the first place. So the answer was near unanimous, and it was the only one this stretch of the protein could ever have produced: the mutation made the receptor fire on its own.

The GS domain mutation that the field had been using as a reagent in labs is Q207D, and that is how it is referred to. The mutation swaps glutamine, an uncharged amino acid, for an aspartate, a negatively charged one, at position 207, right next to the FOP mutation. The mutation was not built with FOP in mind, and it could not have been. It was built years before anyone knew the disease had anything to do with this gene, for an ordinary reason. If you want to know what a cell does when this pathway is running, the simplest way is to hand the cell a version of the receptor that is running all the time, and watch. Q207D was that version, made by copying onto this receptor the trick that had already worked on its relative.

The FOP mutation, R206H, sits one residue away from the reagent Q207D, side by side in the same short box. So when the FOP mutation was published in 2006, it arrived one letter away from a spot that the field had been deliberately mutating for years to force the receptor constitutively on. So, the disease mutation looked like the lab's own switch, made by nature.

The amino acid swaps in the two adjacent mutations though are not the same. The amino acid swap in Q207D adds negative charge, mimicking the effect of phosphorylation. That is why Q207D makes the receptor behave as though the type II kinase has already fired at it. R206H is different: it puts no negative charge, and it imitates no phosphate. The 2006 paper's own modelling claimed only that the amino acid change makes a short helix less stable, and this might loosen the guard protein's grip. Same neighbourhood, different logic.

Any residual doubt the field had about reading the FOP mutation's mechanism from Q207D was cleared when a real FOP patient5 with a mutation on exactly the same residue as Q207D arrived in 2009. The patient carried the Q207E mutation, which swapped the same glutamine but this time for a glutamate (instead of aspartate in Q207D). Glutamate is also acidic, like aspartate. A natural version of the Q207D mutation, in a FOP patient, solidified the hypothesis that the classic FOP mutation made a constitutively active receptor.

A standard way to understand what a mutation does to a protein is to go to the protein's three dimensional structure, and you look at where the changed amino acid sits and what it is touching. The field had such a picture, but it was not of ACVR1. No structure of ACVR1 existed then, and none would for years. What they had was the crystal structure of a relative protein from the other branch of the family, solved in 1999, and it covered only the half of the protein that hangs inside the cell. The half that binds the ligand was not in it.

Everything the field knew about the brake came from that one picture: the box folded back against the kinase and holding it shut, the small guard protein sitting on top of it, the phosphates that the folded shape cannot accommodate so that it has to let go, activating the kinase. Infusing this structural insight into ACVR1 was reasonable as inside the cell these receptors are built alike and share the mechanics.

Everything looks good until you go back to the original structure reference paper and see what it says about the Q207D reagent mutation. It says explicitly that the residue the Q207D reagent copies sits well away from the folded loop and working centre of the kinase, and that its role in switching the receptor on is not apparent from the structure. So, the structure didn't explain even the reagent mutation, let alone the FOP one. Yet, the field anchored its hypothesis on this reference.

In 2008, the constitutive activation hypothesis got another layer of support from experimental data. The FOP mutation was for the first time observed in the living cells, connective tissue cells from FOP patients cultured in a dish. The cells signalled with no ligand added to them. Emphasis on the term 'added', as we will come back to this experiment soon. So the hypothesis was no longer a guess but a real one with empirical support.

Then six years later somebody put the receptors carrying the three mutations, the reagent Q207D and the two FOP mutations R206H and Q207E, side by side and tested them. The reagent mutation turned out to be far more active than the other two. The Q207E that shared the same residue as the reagent behaved more like the classic FOP mutation than the reagent one. And of the three, only the engineered one signalled with nothing supplied to it. The authors wrote that the reagent has severe limitations as a model for FOP. That was the first hit to the sand castle that the field was building on the foundation of a reagent mutation for the past eight years. But the real blow that would shatter the castle came a year later from a company that was not working on anything related to this question at all.

The hypothesis turned out not only to have cracks all over it, but also holes. The field either turned a blind eye to them, or acknowledged them and went on holding the hypothesis anyway. A receptor that fires on its own fires wherever it sits. Every cell in the body carries the gene for this one, from conception. So the model would predict bone forming in many places at once and never stopping. However, in the real world, the disease is quiet for years, then one flare, in one place, usually after an injury, then quiet again. It is there in the earliest record of the disease, written down by Freke in 1740: a boy of healthy look, whose swellings had begun three years before he was seen at fourteen.

Then there is a second hole, and it is nowhere near bone. ACVR1 has another job, in the liver, where it drives the production of hepcidin, the hormone that sets how much iron the body takes in. Push that pathway up and iron absorption falls, which is why people born without working copies of TMPRSS6, a brake on that same pathway, are anaemic for life. A receptor that is on all the time would have made FOP a disease of chronic anaemia as well as a disease of bone. Chronic anaemia is not a clinical feature of FOP.

The mechanism

Aris Economides from Regeneron had been standing near this disease for years when the hypothesis took its final shape. He entered the room in 2008 as part of the jury when noggin was under trial. Noggin had been his subject since he arrived at Regeneron in 1993, and noggin works by taking a signal out of circulation from outside the cell, never by touching a receptor. That was the kind of thing he knew how to do.

After ACVR1 was discovered to be the FOP-causing gene, Regeneron decided to go after it with an antibody. The gene appeared as a target of the right shape: a receptor sitting in the membrane with a piece of itself hanging outside the cell, washed by the fluid outside, and that is something antibodies travelling in the blood can grip. But the hypothesis the field was holding on to said the receptor fires by itself, with nothing bound to it. If that was the case, there is nothing a Regeneron antibody from outside the cell can do, and the programme is dead before it started.

So Regeneron was left with no choice but to test the hypothesis that had stood since the gene was found, and that nobody had needed to check, because until then nothing depended on the answer. And they would not do it in a dish, because nobody knew which cell in the body makes the heterotopic bone, and an assay has to be run in some cell. Pick the wrong cell, and a clean answer would mean nothing. So, it was clear. The experiment had to run in a living animal.

A living animal with FOP in every cell could not exist. Put the mutation into a mouse and the mouse dies at birth. The same single letter that people live with for decades is lethal to a mouse. So the experiment needed a different animal: one in which the mutation appears only after the mouse is grown.

Regeneron built it. The mutated stretch of the gene was installed backwards on the opposite strand, where it cannot be read at all. And a normal copy of the gene was put in its place so that the animal develops as an ordinary mouse. Alongside it is placed a Cre-lox system, a standard piece of mouse engineering that cuts and flips marked pieces of DNA on command. Give a grown mouse a dose of tamoxifen and the switch is thrown: the normal copy comes out, the mutated stretch turns around in the right direction, and from that moment the animal is reading the mutated version of the gene.

That solved two problems at once. First, the breeding animals never have the switch thrown, so they remain healthy and propagate the colony. Second, the animals that are switched on are adults, and they behave like patients. Injure a muscle, bone forms in it.

Now that the FOP animal model existed, Regeneron asked the question the field had gone nine years without settling: does the FOP receptor need a ligand to fire? One way to test it is to take away the ligand and see whether the disease still happens. Rather than guessing which ligand to take away, Regeneron just mopped every ligand this receptor could be answering to out of the tissue using a decoy, which was the ligand-binding part of a type II receptor fused to an antibody tail. It floats in the space between cells and catches ligands before they reach a real receptor, and never the receptor itself, which is left whole, kinase and mutation included. To cover all the ligands, they used the broadest decoys they had, ACVR2A and ACVR2B, broad enough that the answer could not escape them, without yet suspecting what the answer was.

Given to the mice before the injury, it blocked heterotopic bone completely. Not less bone. Just none. A receptor that fires on its own fires whether or not anything is there to bind it, so emptying the space around it should have changed nothing at all.

The animal experiment proved that the FOP receptor does not fire on its own, and that it needs a ligand to switch it on. Regeneron then settled the evidence from the past, the patient cells in a dish that had signalled when no ligand was added. No ligand added is not the same as no ligand present. It turns out cells secrete ligands of their own into the medium, and the serum they are fed brings more. The hypothesis that had survived nine years despite the cracks and the holes was finally shattered, by a company that had entered the FOP field through a side door.

The catch-all trap had answered one question and produced another. Having proved that the receptor needs a ligand after all to fire, the next question was which ligand. Two things pointed at activin A. The first is that unlike BMPs6, activin A is not a resident protein sitting quietly in tissue. It is released at damage, by inflammation and repair, which is exactly when and where FOP makes bone. The second was it is right there in the receptor's name: activin A receptor, type I (ACVR1), treated as a misnomer for the two decades since 1993. The answer to the question had been hiding in plain sight.

Two experiments in the FOP mouse settled the answer. Activin A soaked on to a scaffold and implanted into muscle turned to bone. In their normal littermates, the implant did nothing at all. Then the reverse: an antibody that mops activin A out of the tissue, given before the injury, fully prevented the bone formation7. Activin A was enough to cause the disease, and blocking it was enough to stop the disease8.

The antibody used in that mouse experiment was REGN2477, also called garetosmab, the very molecule the FDA approved eleven years later. The antibody was never made for this purpose. It came from a muscle programme at Regeneron that had nothing to do with bone, and it was brought in as a reagent for the experiment, which then turned into a drug.

So, a single letter change in the ACVR1 sequence made the receptor answer a ligand it normally does not. The mutation did not take away anything from the receptor. It still answers BMP exactly as it always did. But it also answers a member of the other branch, a stranger to bone signalling. Note, activin always binds the normal receptor, and that is how ACVR1 got its name in the first place. It just does not fire the signal inside the cell, and that is the new power the receptor received from the mutation.

Why should a ligand from one branch of the family fit a receptor from the other in the first place? Because of the homology the whole family was built on. The parts are variations of a single blueprint, so they still dock across the branch line. What was missing was permission to pass the signal on, and that permission is held in the very stretch where the FOP mutation sits. Tweaking it miswires the two branches of the family.

The drug

When the right target landed, things picked up speed. What normally follows a successful identification of a target is the lead discovery programme. Screen candidates, pick a lead, optimise it, and then some years later something is ready to go into humans. None of that happened here. The antibody already existed, from a muscle programme that had nothing to do with bone, and the experiment that identified activin A had used that same molecule.

So there was nothing left to design. The molecule that discovered the target is the one that went forward against that same target, on the same logic, read out on the same thing. Phase 1 opened in healthy volunteers in July 2016, the first FOP patient was dosed in February 2018. Under three years from a mouse to a human.

The only thing that changed after that was the name: REGN2477 to garetosmab (brand name: Pasatru).

Note, Regeneron's original plan was different. It came into FOP with the intention to develop an antibody against ACVR1 itself, and the whole ligand question had been forced by that intention. Now that the receptor was known to need a ligand to fire, there was every reason to go back to the original plan. In fact, several groups built antibodies against ACVR1, including Regeneron. But then came a surprise.

In cells the ACVR1 antibodies did what they were designed to do. They sat on the receptor and shut the signal down, on the normal receptor and FOP receptor alike.

In the mice they did the opposite. The treated animals made more heterotopic bone than the untreated ones. And they went on making it even when activin A was blocked, which should have left nothing to drive the receptor at all. It turned out the antibody had not just failed to block the receptor. It had taken over the ligand's role.

The fault turned out to be the shape of the molecule, not the part of the receptor it latches on to. A bivalent antibody has two arms, which grip two receptors at once, bringing them together. And for the FOP receptor being held together is itself the activating event. No ligand required. Make the antibody with a single arm and it blocks the receptor exactly as intended. That is a harder molecule to turn into a medicine, and there was no reason to take it on, because the antibody against the ligand was already working.

The surprise also settled something about how these receptors sit on a cell. If a ligand had to gather four receptors into a complex, then an antibody clamping two type I receptors together would achieve nothing, because the type II receptors that trigger the firing would still be away. It appeared that they were not away. The type I and II receptors are already together before any ligand arrives. The ligand does not build the complex. It simply rearranges the complex that is already assembled.

This new insight makes the field's old answer look better now than it has appeared before. So, the receptor was never idle. It was always assembled, always poised, but held shut. R206H did not switch it on by itself. But it lowered the bar for what counts as a legitimate nudge. Activin A clears that bar. So does a Y-shaped protein that happens to grab two receptors at once.

A receptor is a family property. Seven type I receptors are built on one plan, the same resemblance that assembled this family in the first place, with the same switch region inside the cell and the same partners reading them. So anything aimed at ACVR1 is aimed at a shape its relatives carry too. That is the standing problem with the small molecules: the kinase you want to shut down looks similar to the kinase on the other receptors that you don't want to disturb.

A receptor is also a machine, and the ACVR1 antibody experiment had just shown what could go wrong when you target a machine. Taking hold of a machine, you can accidentally start it instead of shutting it down. That is a risk you take on the moment you aim at the receptor.

A ligand is neither. It is cargo. Clamp cargo and nothing fires, because there is nothing in it to fire. A cargo is the part of the system that is not shared. The mutation of course sits in the receptor that's everywhere, but what makes the disease appear in a specific place after an injury is a single molecule that turns up at damage. That is the part a drug can aim at without hitting anything else.

None of this is a new idea. Noggin does exactly this. It never touches a receptor. It wraps the BMP itself and hides the two surfaces the receptors would have to reach, so nothing ever fires. Sequestration is the whole of it.

And there is a last turn in that. Noggin holds down BMPs, and only BMPs. It does not touch activin A. So the protein the field chased for seven years as the cause of FOP could never have treated it. What carried across was the principle: take the ligand out of circulation and leave the machinery alone.

Looking back it is hard to miss. The man who had spent fifteen years taking BMP ligands out of circulation walked into a receptor disease and solved it by taking a ligand out of circulation.

The first real test in patients was LUMINA-1. A phase 2 trial in 44 adults with FOP, garetosmab against placebo for twenty eight weeks. The first patient was dosed in February 2018.

It missed. The primary endpoint was the change in total lesion activity on a PET scan, meaning every active site in the body added together, the heterotopic bone a patient already had when they enrolled in the trial plus anything new they developed after the enrolment. Garetosmab came out 24.6% below placebo, and the p value was 0.07.

The endpoint was counting bone the drug was never going to remove. Garetosmab stops new bone from forming, it does not dissolve bone that is already there. It was something the mice experiments had shown before. Dose the animals three weeks after the injury, formed lesions stopped growing and new ones stopped appearing, but only a few of the smallest ones shrank.

Separate the two kinds of bone and the drug effect appears. Counting only lesions that were not there at the start, the PET scans found 29 in the placebo group and 3 in the treated group. The crossover made the same point: the placebo patients who carried on into the second period had produced 23 new lesions between them, and just one after they moved to garetosmab.

It was a post-hoc analysis, after the primary endpoint failure. It cannot rescue the completed trial, but it presented the clearest evidence to support a fresh trial.

The phase 3 trial, OPTIMA, was launched with the revised primary endpoint. The PET activity measure was dropped altogether and instead a simpler method was chosen: a whole-body CT scan, one at the start and one at the end, and count the lesions that were not there at the start.

63 patients, 56 weeks, placebo against two doses. The placebo group developed 19 new lesions, the high dose group (the dose that was eventually approved) produced two. That's a 90% reduction9, strikingly similar to the post-hoc result from the LUMINA-1 trial. The drug had produced the same result twice. Regeneron also reported a post-hoc analysis counting lesion volume rather than lesion count, which showed a 99% reduction, the number quoted in the press-release.

OPTIMA has still not been published, so every number above comes from the company's release and the FDA's approval announcement.

The treatment comes with a trade-off, which is the consequence of cutting activin A's ordinary functions. Patients lose their eyebrows and eyelashes. They get acne, boils, infected nail folds, extra hair growth in other places, mouth ulcers and nosebleeds. The infections in that list are not a cosmetic matter. There were more of them at the higher dose, and they are printed on the label as a warning, because an abscess or a cellulitis can need treating in hospital. And the drug comes with a second warning, which is the oldest fact about activin A in this story arriving last. Nobody should be on this drug while pregnant, because activin A is a reproductive hormone, and that is why the first trial of the molecule in humans enrolled only women who could not become pregnant.

The approval is only for adults for now, because the trials enrolled only adults. However, the disease does not wait for adulthood. Whatever has formed before a patient turns eighteen and becomes eligible for the drug is permanent. So the version that would come closest to a real cure is the one yet to be demonstrated in a trial: give it in childhood, before the first flare, while the child still has everything to keep.

Some thoughts

Everyone says drug development is hard. It is one of those sentences repeated so often that it has stopped carrying any meaning. What is rarer is to hear someone say which part is hard. There are many hard parts, the money, the manufacturing, the trials, the regulators. But you know what, the hardest of them, in my personal opinion, is the biology.

Evolution has had hundreds of thousands and sometimes millions of years to get a job done, often perfectly. Then something goes wrong inside one of those processes, and a person falls ill, and we arrive with a mere couple of decades of studies and a great deal of confidence about what the fault is and how to fix it. It is almost never that simple.

Understanding a piece of biology takes more than one person, more than one field, and more than one period of time. Parts have to come together before the biology decides to reveal itself. The story I have just told is an unusually clean example of that, and let's retrace its shape one last time.

Start with the receptor, because the biology in this story turns out to be a decent picture of the story itself.

The complex was never taken apart and put back together. It sat on the surface of the cell already assembled, all components in place, poised but held shut. Nothing was missing except the permission for the signal to pass, and the mutation supplied that by lowering the bar for what counted as a legitimate nudge.

The history has the same shape. By the time anyone could explain FOP, every part of the explanation had been lying in the literature for years. Activin A was isolated in 1986. ACVR1 was cloned in 1993 and named, correctly as it turned out, after activin. The BMP pathway was worked out through the 1990s. The families had been assembled and their DNA banked. Not one piece was missing.

So what was the ligand here? I think it was a person.

Aris did not start with a missing fact. He brought a position. He had been standing near the disease since 2008, and he arrived there through the wrong gene, as the noggin expert on a jury. He was inside a company that could make an antibody and build an engineered mouse. He happened to be at the right place at the right time in the right position. The assembled puzzle pieces were waiting for someone standing where he stood to see the shape.

So what to take from all of this?

Much of the work that took decades to solve could be solved in an evening today. For example sequencing of a handful of patients would have solved the variant discovery or some of the receptor or ligand biology would have fallen out of a high-throughput screen. That is all true, but it is not the take-home.

The take-home is the shape of the puzzle and what it took to solve it. I don't think that will change in time. Today's hard problem is some other thing that today's technology cannot touch, and in ten years it will look silly that people fiddled so long with what by then looks like trivial biology. How hard a problem is depends entirely on from where you look at it. What remains the same is the concept of the problem and what solving it demands.

And what it demands is never one thing. It is not a single bottleneck. It is not about just missing data or a missing understanding. Many things have to arrive together, and sometimes the problem itself remains invisible until it becomes obvious for someone to acknowledge it.

There is a saying that fits this. When it is time, it will happen. Until then it will not, however hard anyone goes at it. I half believe that. This story gives the counterexample. The field had the right pathway for ten years and spent them searching the wrong genes. Everything needed was already there and nothing happened. Readiness on its own is not enough. Someone has to be standing there when the time comes.

Which is why I think serendipity and coincidence are in the blood of biological problem solving, and I do not think they can be taken out of it.

The question in the air is whether AI will solve biology and hand us the cures. I think biology is humbling, and that statement is about the subject rather than the tools. Plenty of bottlenecks will be solved by brute-force compute, including the ones this story ran into. Structure prediction, screening, even reading a literature that no one person can hold all in their head, all of that is real and it will help.

But go back in this story to those ten years. There was no shortage of effort or hands. If you had given the field a thousand times the person-hours, you would have got a thousand times as many searches of the wrong genes, finished sooner. The measurement they were building on was correct. The pathway was correct. The node was wrong, and nothing inside the work they were doing was going to tell them so.

That is the part I am not sure if one can automate. Intelligent machines multiply whatever you are already doing. They do not tell you that what you are doing is aimed at the wrong place. And the step that actually solved the FOP puzzle was not a search at all. Nobody solved FOP by looking harder. What it needed was somebody able to take a step back and notice that the frame through which the problem was being looked at was wrong. That comes out of plain curiosity and from coincidence putting a curious person in front of something that needs solving. Ambition, money, a prize, a wish to do good, all of those can be drivers. I do not think any of them is the engine.

One last thing, about me.

I have no stake in this disease. No one I know personally has it, and none of my research work has been on this topic. I worked at a company where this was happening and I knew the person who drove this programme as a colleague with whom I interacted only occasionally. And that coincidence pulled me into this topic.

I have spent more time on this piece than on anything that I have written. And it was out of nothing but curiosity. I am glad I did. I am leaving it with something that I did not have before: a view of a drug programme from end to end, and a feel for how a biological problem actually gets taken apart. You get that not from reading a paper, but following a question until you find all the answers.

I intend to do more of these on GWAS stories. Long form writing teaches me far more than short form does. It also forces me out of the small pond I have been swimming in, and makes me look at the world beyond it, so I can put what I know and what I do not know in a better context.

I wrote at the start that I wish I had done this homework while I was still there at Regeneron, so I could have got a first-hand account of the story from the people who drove the programme. But come to think of it, all those years the whole story was sitting right beside me, and I never took notice. It took a nudge, a press release on the FDA approval, to finally make me look.

A new preprint from Soumya Raychaudhuri's lab has now ripped apart this locus, plucking one base at a time, and arrives at an impressive list of insights. The authors pinpoint exact base that drives the genetic signal, its effect on the protein, leaving the transcription undisturbed, the cell type the effect lives in, and which cell state within that cell type. It's been a while a functional follow up of a common variant GWAS locus felt so satisfying to read.

The CITE-seq datasets and what the authors were trying to do

The starting point of the journey is CITE-seq, and the authors use it for something the technique was not built for. In CITE-seq you tag cells with antibodies that carry a short DNA barcode, then sequence the barcode along with the entire transcriptome, so every single cell comes back with its genome-wide gene expression and a count for each surface protein tagged on its surface. Almost everyone who used this technique has used the protein counts the way you use a flow cytometry gate, to say what kind of cell this is. The protein counts were designed to be mere labels.

The authors here however treated the counts as biological measurements. If you have the same panel across hundreds of genotyped individuals, the count for each protein on each cell type becomes a quantitative trait that you can correlate with genetics like any other trait. The authors pooled four CITE-seq datasets from three published studies comprising more than a million peripheral blood cells from 356 individuals and mapped genetic associations to 148 surface proteins across eight cell types.

The reason for going after surface proteins is that they are what defines a cell state in the first place. A B cell announces what it is currently doing on its own surface, and other cells pick up that message and react. Most of the published protein QTL studies used serum or bulk tissue, and they inform nothing about the specific cell type the protein came from. CITE-seq offered a unique opportunity for the authors to study the genetic effects on proteins that sit on specific cell types.

What they found, and why they picked CD40

Their analysis revealed ~200 surface-protein QTLs (spQTLs) across 78 proteins, of which 19 proteins share genetic signals with an autoimmune GWAS locus with the same variant driving both the disease risk and the protein level change. Nineteen proteins is nineteen possible stories, and CD40 was one of them.

CD40 piqued the authors' interest as it stood out as the most clinically relevant protein. It's expressed on the B cell surface and it drives memory formation, proliferation and class switching. It binds to the CD40 ligand (CD40L) expressed on T cells, establishing the handshake between B and T cells that marks the activation of B cells and further differentiation. CD40 targeting is already in clinical trials.

The CD40 signal was among the strongest. The lead variant is rs1883832; each copy of its T allele tracks with roughly 20% less CD40 protein on B cells. The effect size is big for a common variant and the locus is the same one mapped in the rheumatoid arthritis GWAS, and also in Graves' disease, multiple sclerosis and inflammatory bowel disease. The interesting part of the finding here is the variant did nothing to the CD40 mRNA. A conventional eQTL study would have missed this signal.

Fine-mapping

Here the authors run into the wall that happens for any GWAS locus. An association locus represents a block of the genome where any of the multiple variants that are inherited together could be the disease culprit. This is not a GWAS-specific weakness. It applies to QTL studies as well. rs1883832 being the lead variant is a statement about the block, not the causal variant.

Statistical fine-mapping helps, but only to some extent. It narrows the block to a list of high confidence variants, but it doesn't break the linkage disequilibrium. Taking the fine-mapping results from their earlier GWAS of rheumatoid arthritis, the authors narrowed the list to ten candidate variants, all highly significant and none of them excluded.

The conventional next move is annotation. The authors overlaid the variants on the regulatory map of B cells, which annotates the places in the genome where chromatin is open. Three variants overlapped, all away from the lead variant. They added a fourth, the lead variant, as it lay inside the CD40 transcript, in the 5' UTR. So four candidates moved forward to the next step, three chosen because a regulatory logic nominated them and one because it sat within the transcript.

Base editing in Daudi to pinpoint the variant

The authors used a brilliant approach to identify the causal variant. They disrupted one variant at a time while holding others constant and checked the outcome. Base editing helped accomplish this easily. In Daudi, an isogenic Burkitt lymphoma cell line derived from B cells, the authors edited one base at a time to measure the changes in CD40 surface protein, along with a positive control where CD40 is fully knocked out. Their readout was CRAFTseq, a method this same group published earlier, which reads the DNA of the edited region, the whole transcriptome and the surface proteins from within each single cell. It helped to turn a challenge of gene editing to their advantage. Editing has a downside, which is it is never complete, and often you end up with a mixture of edited and unedited cells. When all measurements (DNA, RNA and protein) happen within each cell separately, this is an advantage. Within an assay, you get natural comparators of different genotypes offering dose-response curves.

What did they find? None of the four candidates moved the CD40 mRNA significantly. And on the protein, one candidate, the one sitting on the transcript, had a significant effect, reducing the protein by about a fifth per copy of the T allele. The usual suspects (open chromatin variants) did nothing to either mRNA or protein.

The mechanism through which the 5' UTR variant directly reduced protein levels is a beauty. The variant sits one base upstream of the start codon, right on the Kozak motif, the short sequence the ribosome uses to recognise where to begin translating. The ribosome loads at one end of the transcript and begins scanning. It doesn't always stop at the first AUG codon it bumps into. It's the Kozak sequence next to AUG that tells the ribosome to stop and begin. Weaken it, and a fraction of the ribosomes scan straight past, producing no protein at all. The transcript is there in normal numbers, it's just that fewer proteins get translated from it. That's why the variant is invisible to transcription but visible to translation.

The mechanism itself is not new. In 2005, Jacobson and colleagues had the same variant in hand, already associated with Graves' disease, and ran an experiment. They placed the DNA template in a tube, a cell-free transcription and translation system, and measured how much CD40 protein was produced. The T allele made 15.5% less than the C allele. They then looked at B cells from people carrying each genotype and saw the same effect scaling with dose, 13.3% less CD40 in CT and 39.4% less in TT compared to CC. They called it translational pathophysiology. What they could not do at that time was pinpoint that this variant is the causal one driving Graves' disease risk. The mechanism and the exact base inferred by Jacobson et al. turned out to be right, but nothing in their experiment told them whether this base, rather than the nine sitting beside it, drove the protein change.

Validations

With the causal variant caught, the authors further validated their results using independent experiments. A larger editing experiment on rs1883832 alone in Daudi reproduced the finding. Then they used human donors with homozygous genotypes for both alleles (2 CC and 1 TT). In their B cells cultured with CD40L, they flipped the allele to T in the CC donors and to C in the TT donor, and observed the effects. Same result: protein down in proportion to the number of T alleles and no effect on mRNA.

Characterizing the trans-effect

Now they had nailed the causal variant. Next they went looking for what else in the transcriptome moved with this variant. As the CRAFTseq dataset from the edited primary B cells already had surface protein expression and the whole transcriptome data, they looked for trans-effects of the CD40 variant. 213 genes and 68 surface proteins were associated. Reassuringly the genes were enriched for CD40 signalling pathways and interestingly, many were established RA GWAS risk genes.

The dataset also let the authors cluster the cells to find subsets that represent different states of B cells. Looking at the trans-effects within each subset, only one subset showed the trans-effects. By contrast, all subsets showed the cis-effect. This suggests that the trans-effect is downstream of CD40 activation and that only a subset of the B cells are in that activated state. Their subsequent experiment supported exactly that.

To test that directly, they had to take away CD40L, the T cell ligand that engages CD40 and is supplied in the culture. That is not possible in primary B cells, which don't survive without it, so the authors went back to Daudi. Comparing across the genotypes in Daudi with and without CD40L, the trans-effects showed up only in the presence of CD40L.

Cell-state mapping

Having now established that only a specific subset of B cells manifests the trans effects, which are downstream of CD40-CD40L signaling, the authors went on to characterize the B cell states that map to this subset. The markers expressed by these cells told the authors that they resembled B cells enriched in the germinal centers, along with T cells holding the CD40L on their surface. The authors used a published cell type atlas based on healthy human tonsils, which contain germinal centers. As expected, >85% of the activated naive B cells mapped to light-zone states in the atlas. So, the CD40 variant is acting through signaling that happens only in B cells sitting in germinal centers, freshly activated through engagement with T cells via the CD40-CD40L handshake.

If so, the authors hypothesized that the trans effects cannot be detected in B cells in the peripheral blood, as such cells receive minimal CD40L stimulation compared to those in lymphoid organs. Running conventional trans-QTL mapping for the variant in B cells from population blood datasets, that's what they found: a solitary hit, with no relationship to the effects they had measured in the edited cells.

But what about the disease tissue? The authors next mapped B cells from inflamed RA synovial tissue onto the same atlas, and a subset of them — the ones annotated as germinal-centre-like — landed on the same light-zone states. The cells that carry the trans-effects are present where the disease happens.

Translational relevance

The authors conclude by asking one last question: the blood is blind to the trans effects, but what can it see? Perhaps the shift in the cell state itself, due to the CD40 reduction, that could be captured in the blood? Their intuition was right. Testing it in population blood data, they found that carriers of the T allele have fewer memory B cells and more naive ones — consistent with an older flow cytometry based GWAS that had linked this variant to IgD+ B cell counts. The finding makes sense, as CD40 is the ticket for B cells to graduate to memory cells. Lower CD40 protein would therefore shift B cell type proportions in the blood.

That leads to the translational insight. B cell depletion is already a tested idea to treat autoimmunity and it works, of course with caveats like infection risks. But the authors here raise an interesting question: do all B cell types need to be lowered? Maybe only the cell states that are relevant for the disease? The authors' work here supports that notion. In fact, treatments based on CD40 and CD40L targets are in trials. Iscalimab (CD40 blocker) works in Graves' disease, frexalimab (CD40L blocker) is in phase 3 for MS.1 But neither CD40 nor its ligand is a state-selective target. Every germinal center runs on CD40, including the one driving the vaccine response, and it is not confined to B cells. The authors' work opens the therapeutic path to find disease-selective B cell states to safely target to treat autoimmune diseases. Perhaps a CD40-like exercise is needed for a few other genes that were part of its trans network, and that would shed further light on the path.

First about the biomarker
TG and HDL move in opposite directions for a single reason, and it is not lipid handling as some might assume. It is how much room the fat tissue has left for storage.

Insulin instructs the adipose tissue to keep the fat stored and keeps sending fat to adipose tissue for storage. It does so by switching on lipoprotein lipase in the circulation and switching off lipolysis in the adipose tissue. When the storage capacity runs out, insulin fails its task and TG stagnates in blood, CETP trades it into HDL, which dumps it in the liver, sheds its apoA-I, which the kidney clears.

Two consequences: TG rises and HDL falls for the same underlying reason. Hence, the ratio says something that neither individually could.

Before you can use a biomarker as a discovery phenotype, you have to show it measures what you think it measures. The authors did that in two tiers.

Cross-sectionally, a higher TG:HDL ratio tracked fasting insulin, visceral fat, liver fat (on imaging) and steatosis grade on biopsy in bariatric surgery patients, HbA1c, blood pressure and CRP (inflammation). None of these are lipid measurements; all measure fat storage and insulin resistance.

Prospectively, among people free of the outcome at baseline, a higher ratio predicted incident type 2 diabetes (T2D), myocardial infarction, fatty liver disease (MASLD), and liver cirrhosis.

So the number hidden in a routine lipid panel is not reporting about your lipids, but how close your fat storage system is to its limit, years before any disease manifests.

Now to the actual study: the ExWAS in more than a million people, a scale that only RGC can reach.

Collapsing rare variants within individual genes and testing their collective effect on the biomarker, the authors identified 59 genes in which rare coding variants significantly altered the TG:HDL ratio. Not surprisingly, many are what you’d expect: 7 in the LPL pathway, 5 apolipoproteins, 11 monogenic lipid disorder genes, 31 known drug targets of which 23 are already approved or in human trials.

Two results validated the phenotype.
1. Enriched pathways were dominated by lipid and glucose regulation, insulin resistance, lipolysis regulation and fat storage.
2. Tissue enrichment pointed strikingly at two organs and nothing else: liver at 13.6 fold and adipose tissues, visceral at 18 fold and subcutaneous at 31 fold (!)

Adipose tissue took the spotlight, showing that the genes found are concentrated in the tissue that stores fat, validating what biology the biomarker was expected to measure.

Now we dive into the list, move aside the boring ones and bring the surprising ones to focus.

Cross off what you already knew. Strike the apolipoproteins, strike the LPL pathway and their regulators, strike CETP, LCAT and the rest of the reverse cholesterol transport machinery, strike the monogenic lipid disorder genes, strike PPARG, strike insulin signaling genes. Every one of them is a positive control. None of them is news.

Now about 19 genes stay. Sort them by effect sizes and one stands out clearly: FNIP1 at an effect size of ~0.5 standard deviations, 3 times anything else left in the residue, and importantly in the protective direction.

This is my reading of the table (not an account of how the authors got there). But, you know, a gene with no lipid biology attached to it, carrying the largest unexplained effect on the list, is where anyone would look.

How the FNIP1 genetic signal relates to two sides of the ratio is a mirror image. It lowers the TG by 0.41 SD and raises the HDL by 0.41 SD. The paper scores every one of the 59 genes on how the signal divides between the two components, and FNIP1 lands at 0.5, a perfect split between TG and HDL.

That balance is why the ratio is the sharpest view of it. FNIP1 clears significance for TG alone and HDL alone, but the ratio gives the strongest signal of the three. Adjust for either and the other shrinks. The finding lives in the pair, not in either number.

The association is driven collectively by predicted loss of function variants (pLOFs) that break the gene in different ways. Individually the variants are extraordinarily rare, often found in only one or two individuals among the million studied, but together they reach a frequency that allows statistical testing with adequate power.

And before this paper, FNIP1 had never appeared in lipid genetics. Not for TG, not for HDL, not for the ratio. Search PubMed for FNIP1 alongside lipid terms and you get three papers: this one, and two studies in livestock.

A closer look at all the FNIP1 associations validates the protective signal and gives a picture that is not about how much fat the carriers have but rather how well their fat is stored. Their BMI is not meaningfully different. What differs is where the fat sits and what it is doing. The visceral to gluteofemoral fat ratio is lower on MRI, waist to hip ratio adjusted for BMI is lower, liver fat is strikingly lower, HbA1c is lower and apoB is lower. Same body weight, better placed fat, quieter liver and better glucose.

Then the disease outcomes. Across ~230k cases and ~260k controls, carriers had 60% lower odds of a composite cardiometabolic outcome (coronary artery disease, T2D, MASLD and cirrhosis), at an odds ratio of 0.39; T2D at 0.58 and MASLD at 0.43. Note, without the statistical power that a quantitative biomarker gave, these disease effects could never have reached genome-wide significance. It’s impossible to identify this gene purely from disease outcomes, at any sample size.

Genetic association so far showed partial loss is protective for cardiometabolic disease. What about complete loss? It causes a Mendelian syndrome characterized by immunodeficiency and hypertrophic cardiomyopathy. FNIP1, it turns out, is expressed not only in liver and adipose tissue, but also across the body including immune and cardiac cells that don’t tolerate complete loss of FNIP1. And this flags a safety signal.

So the authors went looking for these signals in the heterozygotes. A composite phenotype covering clinical features of the syndrome showed no association, confirming that the disease effects are recessive and might not be a concern in partial loss.

So, what does FNIP1 encode, and what does it have to do with the TG:HDL ratio? FNIP1 encodes folliculin-interacting protein 1. As the name tells you it doesn't work alone. It binds folliculin (gene: FLCN), and the complex sits downstream of AMPK holding two transcription factors TFEB and TFE3, which drive mitochondrial and lysosomal gene programmes. The folliculin complex acts as a brake that holds the downstream TFs on a leash not letting them inside the nucleus. Lose folliculin or its interacting protein, the TFs enter the nucleus, the cell builds more mitochondria and burns energy.

The genetics of FNIP1’s partner FLCN also align with biology: rare damaging FLCN variants move TG:HDL ratio in the protective direction, though at a weaker strength. Although both seem to show similar protective genetic effects, one wouldn’t target FLCN as, unlike FNIP1, FLCN mutations cause a dominant disease, Birt-Hogg-Dubé syndrome, characterized by pulmonary cysts, pneumothorax and predisposition to tumours, particularly of the kidney. In their dataset, the authors found carriers of FLCN pLOFs had an 18-fold higher risk of pneumothorax and 9-fold higher risk of kidney cancer. It’s interesting to note two proteins of the same complex, having same protective effect, show completely different safety profile. Clearly, FNIP1 is the target worth going after.

The authors did functional follow up using in vitro and mice experiments, which revealed something interesting.

In human liver cells, FNIP1 silencing worked as expected, activating lysosomal and lipid breakdown genes. Knock down FNIP1, brake comes off and cells start burning.

The mice are where it got interesting. Deleting Fnip1 in the liver did nothing. But deleting Flcn or Fnip1 and Fnip2 together protected the animals against weight and fat gain on high fat high fructose diet with proportionally more lean mass. So, the pathway behaved as expected in mice, but the exact target disagreed with humans. It appears in mice Fnip2 (the paralog of Fnip1) covers for Fnip1 causing a redundancy, which is not the case in humans, highlighting again the challenges of using mice to find drug targets.

Rest of the mouse phenotype is what you would want: better insulin tolerance, lower circulating insulin, less liver fat. And in adipose tissue, where folliculin pathway is untouched, the authors show activation of lipolysis and mitochondrial genes. It seems cutting the brake in liver alone is sufficient to drive the changes in adipose tissue too.

So, this is the part that makes drug companies like Regeneron invest in human genetics. A rare loss-of-function variant that protects people is the closest thing we have to a phase 1 trial. The carriers have lived their whole lives with one broken copy of FNIP1. What that resulted in is a better lipid profile, less liver fat, better glucose and around 60% lower odds of cardiometabolic disease.

How would you target FNIP1 that is expressed everywhere and sits inside the cell? FNIP1 is expressed in liver, and mouse work shows hitting this pathway only in the liver is enough to move the whole system. And the liver is a tissue well accessed by successful siRNA medicines. So, a liver specific knockdown of FNIP1 gene expression is the way to go.

What would such a drug do that current ones don’t? It would not be a weight loss drug. But it would be a drug that changes where the body puts its fat and how much of it the liver burns. This is a different lever from anything in the clinic now, and it came from reading a ratio that has been sitting on every lipid panel for forty years.

Original paper:
Hindy, Adam, et al. Nature 2026
https://www.nature.com/articles/s41586-026-10864-2

With the rapidly advancing AI field, I'd be lying if I said I haven't been using it to read, write and think about science. I use AI tools more than you'd guess. Among the many things I've been experimenting with using AI is this new newsletter, GWAS Signals. It's honestly a mere sneak peek into the papers I look at, for the readers who already follow the ones I write about. It's not a flood of everything that crosses my screen, but the carved-out fraction that stood out, the papers that felt worth a few seconds to glance at and see what they are about.

GWAS Signals

A weekly digest of human genetics literature worth reading. Curated by an AI pipeline I built, filtered by my interests in GWAS, rare variants, and drug targets. The opinions about what matters are mine, the curation is automated.

By Veera M. Rajagopal

Like most of you, I keep marvelling at how radically the way we consume knowledge is changing. In the early days, my source for new research papers was mostly Twitter. The little bird told me about all the exciting papers that scientists shared online with real enthusiasm, the work they had just published or preprinted or were about to present at a conference. Once in a while I would check the home pages of Nature, Nature Genetics and a few others. Then came the exodus, when Twitter became X and researchers left in droves. I stayed, but I soon realized that scrolling X was no longer keeping me current with my field. So I went back to building my own knowledge source: RSS feeds for my favourite journals, preprints, subjects and PubMed searches. And then AI arrived. With a little bit of context, I found it could triage my feed remarkably well, and not just triage but summarize the key findings. Now I spend my energy brushing through those findings instead of hunting for the papers, and it helps with my FOMO that comes from knowing the field moves faster than I can read.

So here is how it actually runs now. I still choose the journals and topics I want to follow. AI agents do the rest: they ingest the feeds, apply my interest filter, and present a curated list of papers on a dashboard that I try to visit every day. I give my feedback on the curation, marking each paper as one of the following: love, like, good to know, or skip. The system takes that feedback, adjusts what it presents, and gets a little better at curation each cycle. Once a week it picks around eight papers, often around a shared theme, and drafts them into a newsletter. I do not write that prose by hand, though I keep nudging it towards more clarity and brevity.

I should be honest about why I do this. This whole ritual is mainly for my own reading. It is an attempt to stay current with a field that moves faster than I can, and I would want to keep doing it even if no one else read a word of it. But putting it out in the open adds a small discipline. Knowing a few of you are on the other side of it makes me a little more faithful to the ritual than I would be on my own. The ritual itself might not remain the same; it will evolve over time based on how the technologies underneath the workflow evolve, but the effort to have a ritual to read the literature will stay. And I'll try to make it useful for you too.

One caveat to note. These write-ups are drafted by AI, and AI can hallucinate. I read the draft, highlight key sentences, and catch errors before I publish. Yet, hallucinations could escape my attention and bleed into the draft. So, treat GWAS Signals as a pointer, not a verdict. When a summary makes you lean in, when a finding catches you enough that you want to lean on it, refer to the original paper. The signal is meant to send you to the source, not to stand in for it.

You can find GWAS Signals at gwassignals.substack.com, or just hit subscribe below.

GWAS Signals

A weekly digest of human genetics literature worth reading. Curated by an AI pipeline I built, filtered by my interests in GWAS, rare variants, and drug targets. The opinions about what matters are mine, the curation is automated.

By Veera M. Rajagopal

Before I share my insights, let me briefly walk through the plot for context. The plot visualizes the journey of a set of 22 nominated candidates, back-calculated from attrition rates needed to reach one approved drug, through the phases of development. The midline spine marks the stages along with cost labels from DiMasi et al. 2016, the industry-standard reference, expressed per approved drug. The upper area curve is the cumulative cost (in 2013 dollars). The lower area curve shows the transition rates based on data from BIO/Informa (2021), Paul et al. (2010) & Waring et al. (2015) Nat Rev Drug Discov.

Cost of failures, not success

The most quoted number in drug development is also the most misunderstood one. When people say it costs billions to develop a drug, they picture a single molecule being shepherded from lab bench to pharmacy shelf at enormous expense. That is not what the number means. The billions are not the cost of one success. They are the cost of failures—all the failures that were necessary to produce that one success. Every candidate that got nominated, tested and quietly abandoned contributed to that figure. The billion-dollar headline is a measure of failures a company must stomach for one success.

The invisible part of the funnel

Most widely discussed failure rates in drug development start the clock at Phase I. That is actually a generous starting point. Before a drug ever touches a human, it survives a brutal pre-clinical filter that never gets a mention. Based on the limited data available, around 40% of formally nominated drug candidates never make it to human trials. The famously quoted “1 in 10” drug success rate does not count the preclinical attrition. If you factor it in, the odds of success drop to 1 in 22.

And even 1 in 22 is still optimistic as it starts counting only from formal nomination. Before that there are phases like target exploration, hit identification and lead optimization. That earlier funnel, from first exploration to nominated candidate, is almost impossible to quantify at an industry level. It lives inside company R&D pipelines and remains proprietary. The nominated candidate is already a survivor before it enters the visible funnel. The true odds are therefore likely worse than 1 in 22.

Pre-clinical costs rival clinical

Clinical trials, especially late-stage, have a reputation for being expensive. DiMasi et al. estimates $255M per candidate entering Phase III versus $59M in Phase II and $25M in Phase I. That steep cliff before Phase III is exactly what makes the “funnel”. Every gate before III exists to prevent quarter-billion-dollar mistakes.

But here is what that framing misses. Pre-clinical development is invisible in most cost discussions, yet in aggregate it is not cheap. DiMasi reports $430M out-of-pocket pre-clinical spend per approved drug, which is an aggregate cost spanning the entire pre-human pipeline. The data does not allow a per-compound breakdown. Now compare that to our portfolio-level trial costs: 7 entering Phase II at $59M each is $413M, 2 entering Phase III at $255M each is $510M. The most expensive phase per trial and the most invisible phase in the pipeline cost roughly the same. And nobody talks about the second one.

The Phase II graveyard

If you look closely at the transition rates, one number will stand out: 29% of drugs from Phase II make it to III, the narrowest part of the funnel. The killer here is not safety, it’s efficacy. Waring et al. found that pre-clinical failures are dominated by toxicology (59%) and Phase I failures by safety signals (25%), which makes sense as we have reasonably good early tools for catching dangerous compounds before they cost too much. But Phase II failures are led by efficacy (35%), because there is no pre-clinical substitute for asking whether a drug actually works in humans at therapeutic doses. That question can only be answered in Phase II, expensively, after millions have already been spent getting there.

The implication here is to invest disproportionately in early efficacy signals not because safety does not matter, it does. But it usually declares itself early. Efficacy ambushes you late during the most expensive phase before III, and by then the bill is already large.

Buying the race, not the winner

We often come across news of billion dollar acquisitions in the biotech field, which might make you wonder how all that we discussed so far applies there. A company that began with just one target successfully navigated their way into late stages of trials and got acquired for billions of dollars. On the surface it might look like one company is being bought for their one success. But that’s not the full story. That company is just one survivor out of dozens if not hundreds of parallel single-target companies that ran a similar race and quietly failed. They never show up at the deal table, but in reality they are all priced in. The buyer is not paying for what that one company spent. It’s also paying for what other failed companies spent in that target space. The truth is the market ran a portfolio experiment across many bets, and this acquisition settles the tab. Whether the winner got there by conviction or pure luck does not matter. What matters is the buyer bought their way to the end of the funnel by paying what it would have cost to run the race themselves across hundreds of candidates.

Not broken. By design.

It is worth stepping back and wondering if the funnel reflects a broken system that needs fixing. Of course, not. The shape you see is not a failure of the system, it is rather a deliberate design of drug development. The logic is front-loading of attrition: fail cheap, fail fast, and invest heavily only in the survivors. Pre-clinical cuts are inexpensive. Phase I cuts are manageable. By the time you reach Phase III and spend a quarter of a billion per compound, make sure your earlier gates have done brutal and honest work.

The funnel is not broken. But its shape does raise an uncomfortable question: are early filters aggressive enough? Every weak candidate that slips through the early gates carries an expensive price tag before it eventually fails anyway. The cost of a leaky funnel is not just the money. It is the time, the patients enrolled in trials for drugs that should not have made it that far, and the opportunity cost of resources not spent on better candidates.

References:

1. Paul SM, Mytelka DS, Dunwiddie CT, Persinger CC, Munos BH, Lindborg SR, Schacht AL. How to improve R&D productivity: the pharmaceutical industry’s grand challenge. Nat Rev Drug Discov. 2010;9(3):203–14. PMID: 20168317.

2. Waring MJ, Arrowsmith J, Leach AR, Leeson PD, Mandrell S, Owen RM, Pairaudeau G, Pennie WD, Pickett SD, Wang J, Wallace O, Weir A. An analysis of the attrition of drug candidates from four major pharmaceutical companies. Nat Rev Drug Discov. 2015;14(7):475–86. PMID: 26091267.

3. DiMasi JA, Grabowski HG, Hansen RW. Innovation in the pharmaceutical industry: new estimates of R&D costs. J Health Econ. 2016;47:20–33. PMID: 26928437.

4. BIO, Informa Pharma Intelligence, QLS Advisors. Clinical Development Success Rates and Contributing Factors 2011–2020. Industry report; February 2021.

It's been a few weeks since I posted on Substack. I was attending the World Congress of Psychiatric Genetics (WCPG) conference in Singapore. The first time I attended WCPG was in 2014 in Copenhagen, Denmark. It was a turning point in my career. I met a group of wonderful researchers from Aarhus University through a friend, and long story short: I quit my job in India and moved to Aarhus, Denmark, in September 2015 to embark on a full time research career. I spent next 5.5 years in Denmark and then moved to New York to work for Regeneron as an industry scientist.

I've been attending WCPG every year for the past 10 years. Time flies! With my current work profile, I am starting to feel distant from WCPG, scientifically speaking. To be honest, the reason for me to continue to attend this conference is to meet my former colleagues at Aarhus University, who are now my close friends. One thing that I envy the Danes most is their love for travel. They never miss an opportunity to travel, explore the world and enjoy life. This year, they planned a week long vacation in Bali, Indonesia. I decided to give myself a break and tag along with my friends. We spent a wonderful 7 days in different parts of Bali exploring its people, food, culture and importantly, its immense natural beauty.

As I've spent a little too much time away from work, and soon will be traveling to India, I've decided to skip the American Society of Human Genetics (ASHG) conference (the one that is most relevant to my line of work) this year. Expectedly, the FOMO is hitting me strongly. So, I decided to browse through the abstracts to learn about the most interesting talks scheduled for this year. During last year's ASHG, I wrote a long Twitter post, highlighting some of the talks that I found interesting. Many found my post useful. So, I decided to do the same this year for the benefit of readers, many of whom I am sure are now in Denver, Colorado, preparing themselves for five days of exciting science.

The range of topics typically presented at the ASHG is exhaustively wide, and it's near impossible to cover them all. So, I am restricting to the ones that are most relevant to my interests. I'll highlight the talks under various themes.

Exomes

As you may know, at Regeneron Genetics Center (RGC), we are mainly focussed on studying rare coding variants in the human genome and their phenotypic associations. RGC, along with six other pharma companies, funded the whole exome sequencing of all half a million participants of the UK Biobank. This data was made available to all researchers across the world, and it has been a few years since the data was released. We published an initial analysis of this dataset two years ago, and so did other industry teams. But I believe that there are much more discoveries and insights waiting to be found in the UK Biobank exomes.

So, one of the things that I specifically look for every year is new genetic discoveries made using UK biobank exomes. One such highlight last year was the discovery of new obesity risk gene BSN, encoding a synaptic protein basoon, with an effect size comparable to Mendelian obesity genes like MC4R. This work done by researchers at MRC Epidemiology Unit in University of Cambridge was published in Nature Genetics early this year. Unfortunately, as far as I can see, I didn't find any similar major discoveries driven by UK Biobank exomes being presented at this year's ASHG. If you find anything, please post in the comments.

Nevertheless, a few exome sequencing studies caught my interest.

• On Friday, Frederick Satterstrom from Broad Institute in Cambridge, USA, is presenting a large exome-wide association study (ExWAS) of autism based on more than 60,000 cases and 170,000 controls, the largest to date. This work was also presented at WCPG by Jack Fu from Broad Institute. This work represents a big progress in autism genetics. The increased sample size has massively boosted the statistical power for gene discovery, tripling the number of genes linked to autism. Still the discovery is being driven primarily by de novo mutations, though we are starting to see more inherited autism risk genes as well.

• There are two presentations based on an exome-wide association study of inflammatory bowel disease (IBD) by Mingrui Yu (Wednesday, oral presentation) and Ruifei Zhu (Wednesday, poster presentation) from Broad Institute based on nearly 40,000 cases and 65,000 controls. I am particularly interested in Yu's talk where the authors report that individuals who carry cystic fibrosis mutations are protected from IBD. The abstract reads "we found that delF508, the predominant CF-causing variant that accounts for 70% of all CFTR mutations observed in CF patients, has a significant protective effect against IBD (p=1.7E-10, beta=-0.30, se=0.048). This association was successfully replicated in the follow-up dataset (p=3.1E-05, beta=-0.16, se=0.038)." It's a fascinating finding, particularly when you learn about the possible biological mechanism the authors hint at in the abstract. "It is shown that CFTR serves as epithelial receptor for S. Typhi transluminal migration and that heterozygous deltaF508 mice translocated significantly fewer S. typhi into the gastrointestinal submucosa than wild-type CFTR mice. Therefore, it is plausible that the protective effect of CFTR in IBD may stem from similar interactions with yet unidentified bacteria."

• On Tuesday's plenary session, Duncan Palmer from Big Data Institute in Oxford, UK, is presenting a large-scale rare variant association analysis from the Biobank Rare Variant analysis (BRaVa) consortium. It is a collaboration across 16 biobanks across the world that has harmonized rare variants identified in more than a million individuals. For more details, check out their website. There is also a poster on Thursday by Frederik Lassen from Oxford University on rare recessive effect association study based on an analysis of compound heterozygous variations in the BraVa dataset.

• Genes and Health Study is a biobank dedicated to British individuals of South Asian origin. This is one of my favorite biobanks and I've highlighted in the past examples of rare human knockouts identified in the Genes and Health cohort:

  • a British-Pakistani woman who was found to be a knockout for HAO1, encoding the a liver enzyme, which is the RNAi target of Alnylam's lumasiran, now an FDA approved medicine for the treatment of a rare disease called primary hyperoxaluria type 1.

  • a British-Bangladeshi man who was found to be a knockout for MC3R, encoding melanocortin 3 receptor, whose biological role in childhood growth and puberty timing came to light only recently through an exome-wide association study of age at menarche in the UK Biobank. Thanks to the Genes and Health participant, who has helped scientists to understand the phenotypic consequence of complete loss of MC3R.

  Many exciting work based on Genes and Health Study biobank are being presented this year at ASHG. On Thursday, Hye In Kim from Pfizer is presenting on an interesting analysis of human knockouts identified in the Genes and Health Study based on exome sequencing of 44,028 British South Asians. A part of this analysis links the genetic insights from rare homozygous loss of function variants with clinical trial success. This is in relation to one of my recent posts on the value of human genetics in predicting drug development success. Here the authors find that drugs that work by inhibiting genes are more likely to succeed, when individuals completely lacking this gene are found to exist in the general population (as evidenced by homozygous loss of function variant carriers in Genes and Health Study). If I were attending ASHG, I'd definitely won’t miss this talk.

Genomes

Recently, UK Biobank has released whole genome sequencing data on its half a million individuals. My last Substack post was on the cost effective value of WGS vs WES in terms of gene discovery. As I have discussed in the post, there are many challenges to address before the field can switch from WES to WGS for gene discovery at population level. So, I am obviously super interested on all the work presented at the ASHG based on large scale WGS datasets. In addition to UK Biobank, academic researchers now also have access to WGS data of more than 150,000 individuals from All of Us biobank. Not surprisingly, this year there are many presentations based on WGS data from UK Biobank and All of Us.

• On Wednesday, Ryan Dhindsa from Baylor College of Medicine in Houston, USA, is presenting an analysis of inherited chromosomally integrated human herpesvirus 6 (HHV6) in more than 730,000 human genomes. This is one of my favorites this year. I've known many use cases of WGS data. But I've never thought about this particular one: identifying individuals who carry a HHV-6 viral genome integrated into their germline. Apparently, HHV-6 is the "only virus known to transmit through the human germline". The authors report that 1.1% of the 730k individuals carried HHV-6 in their genomes and these individuals had an increased risk of skin cancer, particularly basal cell carcinoma. It's previously known that HHV-6 viral DNA is often detected in the basal cell carcinoma tumors. The new finding suggests that "germline, rather than somatic viral exposure, predisposes individuals to basal cell carcinoma". Fascinating!

• On Thursday, Konrad Karczewski from Massachusetts General Hospital, USA (gnomAD team) presents an all x all common and rare variant association analysis in 245,000 whole genomes from All of Us. This team from Broad Institute is reputed for their work in building useful genomic resources like gnomAD browser, genebass, pan-UKB GWAS downloads etc. The authors plan to meta-analyze results from UK Biobank and All of Us and release data iteratively with a final target sample size of 1 million whole genomes.

• One of the problems of analyzing rare variants from noncoding genome as I've discussed in my previous post is the lack of well defined genomic boundaries like we have for gene sequences and the lack of well characterized variant effects like we have for coding variants. However, resources like ENCODE do provide a map of promoter and enhancer regions in the human genome across a range of cell types as a starting point to perform aggregate rare variant associations. On Friday, Jack Flanagan from Seoul National University in Republic of Korea presents on region-based rare variant analysis of UK Biobank whole genomes. The abstract reads "By leveraging data on enhancer/promoter-gene interactions and key epigenetic markers across over 1,500 cell types, our analysis provides a better understanding of the role of rare variants in complex traits."

• On Friday, Harry Wright from University of Exeter in Exeter, UK, presents a WGS-based rare variant association analysis of anthropometric traits in 750,000 individuals. The authors have discovered some interesting new rare variant associations, for example, a 5' UTR variant near FGF18 having a large effect association with height. The authors highlight that this variant is loss of function intolerant and that is why earlier exome sequencing-based analyses failed to find an association with height. The same team has recently published a WGS-based rare variant association of height in 333,000 individuals.

• Apart from All of Us, an another biobank that contain WGS data is TOPMed (though not at the scale seen in UK Biobank and All of Us). On Wednesday, Margaret Sunitha Selvaraj from Massachusetts General Hospital in Boston, USA presents a WGS-based noncoding rare variant association analysis of LDL cholesterol in 246,000 individuals from UK Biobank and TOPMed cohorts.

• One other obvious use case of WGS data is estimate heritability of complex traits based on both genome-wide common and rare variants. The Twitter- and Substack-famous Sasha Gusev (highly recommend Sasha’s Subtack) has asked multiple times on Twitter why no one is doing heritability analysis using UK Biobank WGS data. Hyein Jung from Kyung Hee University in Seoul, Republic of Korea has now answered Sasha's request. Jung is presenting on Saturday on rare variant heritability analysis of complex traits using WGS data from the UK Biobank. The authors report that total heritability of height based on common and rare variants is whopping 82.25%, which is almost all of the twin heritability, which ranges between 80 to 90%. The abstract reads "Our results showed that for height, we accounted for 82.25% of the 90% twin heritability, while for BMI, we accounted for 39.37% of the 90% twin heritability." It seems the missing heritability of height is found, but not that of BMI. I wonder why we couldn't find for BMI. What's happening with the BMI heritability?

  

Structural variants

One important use case of WGS is study structural variants (SVs), particularly the rare noncoding ones. This year, there are many talks on this topic.

• On Friday, Santosh Atanur from AstraZeneca in Cambridge, UK, is presenting a phenome-wide association study of structural variants identified in 460k UK Biobank genomes. The authors write in the abstract that 98% of the SVs are noncoding, of which 11% spanning known enhancer regions. The authors also highlight few interesting examples such as a 5kb deletion overlapping a cardiac pericyte and vascular smooth muscle enhancer increasing the risk of atherosclerotic heart disease.

• On Wednesday, Simone Rubinacci from Brigham and Women's Hospital in Boston, USA, also presents an SV analysis based on 500k UK Biobank genomes. The authors have discovered many interesting associations including an Alu insertion in the promoter of an endothelin gene (EDN3) associated with blood pressure, which reminds me of the famous PHACTR1 saga.

• On Wednesday, Emma Pierce-Hoffman from Broad Institute in Cambridge, USA, is presenting on structural variant discovery in ~100k All of Us whole genomes.

• Once important benefit of having SV data for hundreds of thousands of individuals is to build the SV constraint map of human genome, that is, to map the critical regions of the human genome that are intolerant to large structural changes. Of course, gnomAD team is on it. Xuefang Zhao from MGH in Boston, USA, is presenting a poster on Thursday on the "Functional impact of 2.7M structural variants across global populations".

• On Wednesday, Shubham Saini from 23andMe, USA, presents analysis of copy number variants (CNVs) identified using genotyping data of 5 million 23andMe participants. Given the sheer scale of the data, the authors were able to identify not only common but also rare CNVs and their phenotypic associations. The research work that comes out of 23andMe have never failed to amaze me. Such a great resource! Yet, the company is not getting a break from bad press recently. I hope the company turns things around soon and gets back on its feet.

Polygenic risk scores

As always, you can find hundreds of presentations on polygenic risk score (PRS) in any human genetics conference. Consequently, the bar to get excited about a work related to PRS keeps rising year after year. Currently, the field is interested more on clinical utility of PRS. Particularly, I am interested in the clinical value of PRS as a screening tool rather than a diagnostic tool. We have seen good examples in the past, for example, screening for individuals at risk of fracture using PRS of bone mineral density. On that front, one abstract caught my eye. On Wednesday, Rosalind Eeles from The Institute of Cancer Research in UK is presenting the results of BARCODE 1 study that evaluated the value of prostate cancer PRS over traditional screening tools (such as PSA) in identifying men in middle to elderly age groups at risk of prostate cancer. The abstract reads "It detects a high proportion of clinically significant disease compared with PSA or MRI based screening programs and MRI missed a significant proportion (17-67%) of cancers found on biopsy. This is the first study to assess if this approach will be useful in population screening programs."

Speaking of PRS, another abstract caught my interest. Many of you might be aware of the resource PGS catalog which hosts PRS weights for hundreds of phenotypes that can be used to generate PRS in your cohort without the need to train models yourself. The Finngen team has used all the models available from PGS catalog to generate PRS for more than 3000 phenotypes in 400,000 Finngen participants and tested associations withs nearly 5000 clinical end points. The results are made available through a "PGS browser". Nikita Kolosov from The Ohio University College of Medicine in USA is presenting this work on Wednesday.

Proteomics

UK Biobank released Olink proteomics data measured in 50,000 participants a year ago. Again, this is an industry-led effort called Pharma Proteomics Project (PPP), and the flagship papers came out in 2023. I've always been excited about the various applications of the PPP resource since the beginning (as you can see in my 2022 round up and Genetics Podcast interview). Among the many applications, one specific application, which I think we will hear more and more about in the near future, that I am closely watching for, is the value of proteomics in predicting disease risks. As far as I've understand, the consensus is that proteomics risk score is orthogonal to polygenic risk score as it is capturing a lot of environmental risks. This is going to prove valuable in drug development, particularly for developing biomarkers to assess clinical trial results. On Wednesday, Manik Garg from AstraZeneca is presenting a disease prediction model called "MILTON" trained based on plasma and urine biomarkers including Olink proteomics data in the UK Biobank. The authors note that addition of proteomics data remarkably boosted the prediction performance. The authors further performed genetic association analysis of MILTON-predicted phenotypes to identify new associations, which they replicate in Finngen data. The authors have also provided the results via a browser.

Metabolomics

Speaking disease prediction using proteomics data, an another layer of omics data that has been recently generated for all half a million UK Biobank participants is plasma metabolomics by company Nightingale Health. On Wednesday, Jeffrey Barrett, the CSO of the company, is presenting on the first pass analysis of this huge dataset. The authors generated metabolomic and polygenic risk scores for 30 chronic diseases and compared the prediction performances between the two. The abstract reads "The metabolomic scores are more strongly associated than polygenic scores for all diseases tested except common cancers, and the metabolomics tracked observed changes in risk profile across time in longitudinal samples". This is reminiscent of what we are seeing using the proteomics risk score. Like, proteomics risk score, I expect that metabolomic risk score too is orthogonal to polygenic risk score. The authors have further performed genetic association analysis of metabolomic phenotypes and found many interesting associations. The authors also note in the abstract that UK Biobank metabolomics dataset and related GWAS summary statistics will be made available to researchers in Autumn 2024.

Drug targets

If you have even a slightest interest in using human genetics to advance drug development, you shouldn't miss the plenary session titled "The Promise and Payoff of Human Genetics and Genomics: Paths from Bench to Bedside" scheduled for Saturday. I am particularly excited about two of the talks, one by David Goldstein (CEO of Actio Biosciences) and the other by David Altshuler (CSO of Vertex Pharmaceuticals), pioneers and reputed leaders in the human genetics field.

Therapeutics

This is one of my favorite themes at the ASHG. Every year we get to hear some fascinating story of innovative drugs being designed to treat challenging rare diseases. Last year, I highlighted a creative therapeutic design to treat Angelman syndrome. This year we have an another interesting story. On Wednesday, Alban Ziegler from Columbia University in New York, USA, is presenting about an intrathecal, allele-specific antisense oligonucleotide therapy designed to treat an individual with a rare disease called KIF1A associated neurological disorder (KAND). Here the allele-specific targeting is achieved by targeting a noncoding variant in phase with the pathogenic mutation. This is one of the ways human genetics is helping with drug development: by identifying genetic markers in cis to allele-specifically target the RNA to treat conditions caused by genetic defects in haploinsufficient genes.

I’ve bookmarked more presentations. But I’ll stop here. I hope you find this curation helpful. Before you go, a quick update. As you may know, I’ve been doing quarterly podcast episodes with Patrick Short on The Genetics Podcast. The latest episode was released recently in which I discuss five interesting human genetics work from the third quarter of 2024 (all of which were covered in my past Substack posts). Do check out!

Finally, I’d like to give a big shout out to Patrick Short for his incredible work in hosting The Genetics Podcast. He has now completed more than 150 episodes! At this year’s ASHG, Patrick is hosting a get-together for all past, present and future guests and listeners of The Genetics Podcast. Sign up via the link or reach out to Patrick directly via Twitter DM, if you’d like to attend the event.

Regular readers of my Substack will know that I am a noncoding genome enthusiast. I get excited every time I read about a new noncoding discovery, and I highlight such findings in the Substack post. However, as an industry scientist working on drug discovery, if I am asked to choose between whole exome sequencing (WES) and whole genome sequencing (WGS), my money would be on WES. My colleagues at Regeneron Genetics Center (RGC) feel the same.

RGC was one of the eight pharmaceutical companies that invested in the whole exome sequencing of 500,000 UK Biobank research participants. But when industry giants later joined hands to throw money on whole genome sequencing the full UK Biobank, RGC opted out. The leadership, which includes pioneers in human genetics, strongly felt that WES provides more bang for the buck than WGS. They even assigned a team to empirically make this case. The results of their study are now published in Nature Genetics. In the study, Gaynor et al. ask how many more genetic associations can be found by WGS compared to WES and arrive at a sobering answer: not that many.

The current standard in the field to perform genome-wide association study (GWAS) is by combining array genotyping with WES. The array genotyping measures a sparse set of variants (~500,000) distributed across the genome, which are then used to impute unmeasured variants (~10M to 15M) using linkage disequilibrium (LD) maps. This approach helps to confidently identify only common variants (minor allele frequency >1%) as imputation accuracy drops at lower allele frequencies due to poor LD. So, researchers go for sequencing technology to measure rare variants. WES helps find rare variants from the coding regions, which span around 1-2% of the genome. So, the combination of common variants across the genome and rare variants across the exome has been the work horse of the field so far. So, the extra information that WGS provides is rare variants from the noncoding genome. Since the noncoding genome comprise more than 98% of the genome, there has been understandably high expectations for potential genetic discoveries from noncoding rare variants. In their analysis, Gaynor et al. specifically evaluated the yield of new genetic signals from noncoding rare variants by making a head-to-head comparison between WES+IMP and WGS.

The authors analyzed the genetic data of 150k UK Biobank research participants. As expected, they over all found more variants using WGS than WES+IMP. While WES+IMP identified 125M variants, WGS identified 600M variants. But the important question is if this 5x more variant yield will translate to 5x more discoveries. When the authors performed GWAS of 100 phenotypes, they found that the yield of genetic associations were similar between the two approaches. This is not surprising because most of the extra variants that WGS capture are rare variants, particularly singletons (i.e., private variants observed in only one individual). Nearly half of the 600M variants identified by WGS are singletons, whereas only 7% of the 125M variants identified by WES are singletons.

To identify genetic associations with rare variants, we need sufficient number of carriers. So, at the single variant level, the singletons and most of the ultra rare variants will not yield any significant genetic associations. In the WES analysis, we circumvent this low carrier count problem through a burden approach. We collapse rare variants across each of the genes into a single group. For example, if there are 20 rare predicted loss of function (pLOF) variants in, let's say, gene A and each of the variants are seen only in one or a few individuals, we pool all the carriers together resulting in 50 to 100 carriers, thereby boosting statistical power to test genetic associations. However, this collapsing technique doesn't work for noncoding variants, at least, not as as straightforward as it is for coding variants. While dealing with coding variants, we have well defined discrete genome boundaries defined by gene sequence and well defined variant classes such as loss of function, missense variants etc. In the case of noncoding variants, we do not have well defined boundaries and variant effect classes. The existing noncoding annotations like promoters, enhancers, 3' and 5' untranslated regions (UTRs) etc. are blurry and highly context specific. These limitations substantially reduce the number of meaningful discoveries that one can make sequencing the noncoding genome. A good example is a recent paper on noncoding rare variant associations with height, published by researchers at the University of Exeter in the UK.

Hawkes et al. analyzed WGS data of more than 300k individuals from UK Biobank, TOPMed and All of Us cohorts and tested the associations of rare coding variants with height. They used both single variant and collapsing approaches for association testing. After excluding known associations, the authors ended up with mere 29 associations. Only a few of these 29 associations implicate any genes. Some of the associations are of course interesting. They include noncoding variants near genes well associated with height, for example, GHRH (growth hormone releasing hormone), and noncoding variants implicating novel genes in the biology of height, for example, HMGA1 (high mobility group A1). But the question is, are these handful of discoveries worth spending money to sequence the whole genome of 330k participants, when you could use the same money to sequence perhaps 3 to 4 times more participants?

In their study, Gaynor et al. demonstrate the disproportionately higher yield of genetic signals with increasing sample size. While the authors did many interesting comparisons, I'll focus here on one finding that signifies the importance of sample size in gene discovery. The authors compared two extreme sample sizes of ~47k and ~470k. When testing genetic associations using these sample sizes, the authors found that a 10-fold increase in sample size yielded approximately 20-fold increase in genetic associations. Coming back to height, my colleagues have been working on an exome-wide association study (ExWAS) of height in > 1 million individuals. The preliminary findings were presented at last year’s American Society of Human Genetics (ASHG) meeting, and updated results are planned to be presented at this year’s ASHG. I recommend checking the presentation to appreciate the power of sample size in rare variant discovery.

I may sound bearish on WGS. But I am not. I believe WGS has an enormous potential in advancing genetic discoveries. But it will take time. As much as I am against the idea of spending money on WGS for drug discovery, I couldn't be more excited about the fact that there are people who have invested generously in whole genome sequencing the entire UK Biobank. The below tweet from Daniel McArthur, a prominent leader in human genetics, summarize how I and many of my colleagues felt.

One thing that I'll closely watch is what new findings emerge from UK Biobank WGS data. There is no doubt that many fascinating discoveries will be made, as illustrated by deCODE Genetics in their flagship paper on the initial analysis of the first 150,000 genomes. I've highlighted in my past posts some of the fascinating discoveries reported in the paper. As people play around with the full UKB WGS dataset, more such findings will trickle down over the next years.

Personally, I am more interested in the potential of noncoding variants in advancing drug development as I've explained in a recent post on the discovery of a noncoding enhancer variant linked to a Mendelian form of heart disease. Particularly, I am curious about the value of noncoding variants to inform consequences of disrupting genes that are constrained for coding variants. Genes that play critical roles during early development are often intolerant to damaging mutations. Hence, we see rarely any carriers of pLOFs or deleterious missense variants in such genes in the general population. I expect noncoding variants that disrupt such genes, perhaps only in specific tissues or during the late stages of development, might surface in the future GWAS based on WGS data. For example, in the height paper by Hawkes et al., the authors write that the new gene, HMGA1, that they have linked to height via noncoding variants has no coding variant association with height. They discuss that this may be because the gene is mutationally constrained. The authors found that among the 330k UK Biobank participants they analyzed, only one carried a pLOF in HMGA1. So, it may be the case that HMGA1's crucial role in early development may have previously obscured its connection to height, which is now being revealed through the study of noncoding variants.

Speaking of WGS, I should mention about its value in the rare disease space. I've repeatedly mentioned this in the past: important noncoding discoveries will often emerge not from analyses of hundreds of thousands of individuals from large biobanks, but from handful of individuals with rare diseases. We've seen many examples so far (HK1, CHD2, KCNB1, ASIP and RNU4-2). Although these examples have exemplified the potential of WGS in rare disease diagnosis, it is challenging to confidently say that WGS is more cost effective than WES for diagnosing rare diseases. Few recent studies have shown mild to moderate higher diagnostic yield for WGS compared to WES. However, given the multiple fold cost difference between WES and WGS, one would argues that one can provide genetic diagnoses to more families by simply opting to choose WES. Perhaps, WGS could be reserved only for families unsolved using WES. But then there are example such as RNU4-2 that clearly make the case that many common causes of neurodevelopmental diseases (NDDs) remain undiscovered because of the long standing bias towards coding genome. Mutations in RNU4-2, encoding a small nuclear RNA involved in gene splicing, were found to explain nearly 0.4% of all NDDs. That is a big number and makes a strong case for WGS for rare disease diagnosis.

As you can see, unlike common diseases, the question of whether WES is more cost effective than WGS is more debatable when it comes to rare diseases. There are valid arguments on both sides. Nevertheless, the field is quickly moving towards WGS in the rare disease space. Furthermore, the cost of WGS is also continuing to drop. So, WGS will become the default choice of sequencing some day. But it’s difficult to predict how far that day is from now.

For today's post I want to revisit one of the old stories, a research work by my colleagues at Regeneron. I wrote about this on Twitter last year when the work was preprinted. I also mentioned it briefly on Substack in the 2023 round up and discussed it on The Genetics Podcast. But now that it is officially published, I thought I should write a short Substack post about it.

I like this work for many reasons. It hits many of my favorite themes:

• Non-European populations-based discoveries

• Genetic convergence between Mendelian and common diseases

• Therapeutic implications of non-European genetic discoveries.

The discovery

My colleagues (Rodriguez-Flores et al.) did an exome-wide association study of stroke in around 75,000 individuals of South-Asian ancestry from the Pakistan Genomics Resource (PGR) and uncovered an important genetic risk factor of stroke among South Asians. A missense variant, p.Arg231Cys, in NOTCH3 that is 30 times enriched among South Asians (MAF=0.58%) compared to Europeans (MAF=0.019%) was found to increase the risk of stroke more than 3-fold.

CADASIL

The gene NOTCH3 is a known Mendelian gene for stroke. Pathogenic missense mutations in NOTCH3 cause an autosomal dominant stroke syndrome called CADASIL (Cerebral Autosomal Dominant Arteriopathy with Subcortical Infarcts and Leukoencephalopathy), discovered in the 1970s. The condition is characterized by early-onset recurrent stroke that gradually destroy the blood vessels in the brain, resulting in motor and sensory deficits, dementia and death, typically before the age of 60 years. It is through the genetic mapping of CADASIL, NOTCH3 was first cloned in humans.

In addition to its history, the genetics of CADASIL is also fascinating. NOTCH3 is a huge gene with 33 exons that code for 2321 amino acids. It encodes a transmembrane receptor expressed in the vascular smooth muscle cells. The receptor needs to be cleaved to release its intracellular domain from the plasma membrane, which then swims through the cytoplasm into the nucleus to activate the transcription of its target genes. The most important part of the protein, however, is its extracellular domain that interacts with the ligand and initiates the NOTCH3 signaling cascade, which involves cleavage by multiple enzymes (including gamma secretase, the same one that cleaves amyloid beta protein).

The extracellular portion of NOTCH3 is unique and made of six cysteine residues repeating again and again 34 times. So far, almost all the CADASIL mutations (including p.Arg231Cys discovered now in South Asians) were found within the extracellular domain, that either add or remove extra cysteines, both of which disrupt the NOTCH3 function. Any change in the cysteine residues in the extracellular domain triggers protein misfolding, resulting in aggregation, which ultimately damages the vessel walls through infarction and inflammation.

Convergence between Mendelian and common disease

The most interesting aspect of the current finding is it signals a convergence between a rare Mendelian disease, affecting a few thousands and a common disease, affecting more than 100 million individuals around the world.

One of the first things we do when reviewing GWAS results is we search for genes that are already known to be involved in the disease, particularly the ones that cause Mendelian diseases. For example, when you run a GWAS of LDL cholesterol, you'll see tall towers rising from near APOB, LDLR, PCSK9 and many other genes, all of which were known to be mutated in familial hypercholesterolemia. But such convergences rarely occur for brain-related conditions.

The most recent GWAS of stroke involved more than a million participants, and yet there was no signal near NOTCH3. But in the current work, only 75k individuals were involved and there was only one signal, which was near NOTCH3. That is one of the benefits of studying diverse populations: what cannot be found in more than a million individuals in one population can be found easily in few thousands in another.

Why is it important to find a convergence between rare, low frequency and common variants? Risk variants with different allele frequencies will often differ in their penetrance, which gives us an opportunity to study the phenotypic consequence of perturbing the gene at severe, moderate and mild levels. And this knowledge of a correlation between gene disruption and phenotypic change can be crucial for drug development.

Therapeutic relevance

The convergence between Mendelian and common disease also has therapeutic implications. CADASIL is a rare disease, impacting around 2 to 4 per 100,000 individuals. One of the challenges of developing drugs for rare diseases is the small target population. The new finding potentially expands the target population for a drug that targets NOTCH3. Stroke is a common disease, particularly in South Asian populations. The current work projects that p.Arg231Cys variant alone can explain 1-2% of the stroke cases in Pakistan and possibly in other South Asian countries, which will translate into hundreds of thousands of cases. So, now there is a good reason for companies to work on drugs targeting NOTCH3.

The one I explained above is one of the underappreciated uses of studying non-European populations. Diversity not only can help discover new risk genes, but it can also expand target populations for known risk genes, which is a major factor that influences investment decisions in the drug development field. APOL1 is a good example. The reasons why nearly 20 companies are currently working on drugs targeting APOL1 is there is a huge target population—hundreds of thousands of Africans and African Americans. Likewise, we may see companies going after NOTCH3 in the future.

Decoding the noncoding genome has been a recurring theme in my past posts. Today, I have another fascinating story under this theme. It's about a noncoding variant underlying a newly discovered rare, Mendelian heart condition characterized by a distinct ECG pattern and sudden cardiac death.

One of the biggest mysteries of human genome that we will see scientists incrementally solve over the next decades is the biological mechanism through which noncoding variants influence disease-risks and trait variations. Solving this mystery not only will bring genetic diagnoses for hundreds of rare diseases (as we've seen in the recent RNU4-2 story) but I believe it will also bring breakthroughs in drug development.

As a scientist working in drug development, I try to understand how genetic variants increase or decrease gene function, and how this information could be used identify potential drug targets. So far, the field has been mainly relying on coding variants to do this. We specifically look for rare coding variants to understand the beneficial and harmful phenotypic consequences of increasing or decreasing the function of a gene.

There are limitations to relying solely on coding variants. For example, most of the variants often decrease the gene function rather than increase it. Of course, there are examples of coding variants that increase the gene function. But they are not as common as loss of function variants, and also, they are neither easily identified nor confidently interpreted. The ones we confidently interpret today are mostly loss of function variants such as frameshift, stop gain variants that truncates the protein and abolish its function.

Another limitation of studying coding variants is that their effects are often omnipresent in the human body across space and time. A person born with one copy of a protein truncating variant in, let's say, gene X will be deficient of that gene since their conception in utero until their death. They will be deficient of the gene in every cell of their body that expresses it and at every phase of their life when the gene is expressed. Of course there are exceptions like certain splicing variants affecting specific transcripts that show tissue-specific expression, but such examples are rare.

Noncoding variants are expected to address many of the limitations of coding variants like the ones I mentioned above, but we are at our infancy in understanding of how sequence changes in the noncoding genome impact functions of neighborhood or distant genes. Our knowledge are slowly evolving with emerging discoveries, particularly the ones made in individuals with rare Mendelian diseases.

Gene often express tissue and development stage specifically. These space and time restrictions are coded in the regulatory elements widespread across the noncoding genome. Mutations within such regulatory elements can cause severe consequences, sometimes more severe than coding mutations, and through discovery of such mutations, scientists learn about the existence of critical noncoding regulatory elements and the mechanisms through they control their target gene expression.

I've highlighted in my past posts many examples of discovery of noncoding variants underlying Mendelian diseases. One of my favorites is the discovery of intronic variants in HK1 that derepress the hexokinase expression in pancreatic beta cells, resulting in uncontrolled insulin secretion and fatal hypoglycemia in infants with congenital hyperinsulinism. Here, the noncoding variants taught scientists about the consequence of expressing hexokinase in beta cells, where they are not supposed to express.

Another example is the discovery of a noncoding structural variant underlying a monogenic form of extreme obesity that rewires the gene expression program of ASIP, resulting in ubiquitous expression of ASIP instead of skin-specific expression. Here, the noncoding variant taught scientists about the consequence of expressing ASIP in brain hypothalamic neurons where they are not supposed to express.

Another example is the discovery of a noncoding variant underlying a neurodevelopmental disorder that abolishes the expression of a long noncoding RNA and as a result, amplifies the expression of nearby neurodevelopmental gene CHD2.

Among the noncoding variants, the ones that cause gain of function effect is of particular interest for me for the reason that I explained before. From past examples, we have learned that noncoding variants can achieve gain of function effect through duplication of a regulatory element (as was the case with ASIP), loss of inhibition of gene repression (as were the cases with HK1 and CHD2). In a recent preprint, a research team from University of Oxford report a new mechanism through which noncoding variant causes a gain of function effect: de novo creation of a cardiomyocyte-specific enhancer.

The story starts 36 years ago when cardiologists at the Copenhagen University Hospital in Denmark encountered a 30 years old patient who incidentally presented with a puzzling ECG feature: persistent ST segment depression1 without any associated cardiac disease. The patient appeared healthy. The doctors followed this index patient for the next 30 years. After remaining asymptomatic for 25 years, the patient developed atrial fibrillation at 55 years of age. Eight years later, he developed ventricular fibrillation and was rescued from sudden cardiac death. Both his children, the doctors found, had similar ECG changes as his. The doctors learned that two of their family relatives have died in the past from sudden cardiac death. While studying this index family, the doctors also encountered a series of other families with similar history of atrial and ventricular arrhythmias associated with ECG changes strikingly similar to that of the index family. They realized that are dealing with a new, genetic cardiac syndrome, which is passed down in the families in an autosomal dominant fashion. They published the case reports in NEJM in 2018.

The current preprint is a genetic follow up of the cardiac syndrome, named as ST depression syndrome (STDS), reported in 2018. Using linkage analysis of the affected families, the authors mapped a locus in chromosome 20 that was shared by two different families. Through further analysis of sequence variants within the chromosome 20 locus, the authors identified the disease haplotype, which turned out to be same in both the families, and also, in another patient from a third family. The authors suspect that all the affected families are distantly related and have inherited the disease variant from a common founder. Through further analysis, the authors narrowed down the disease variants to a single heterozygous noncoding variant that fell within the linkage region and was found in every affected family member but not in the unaffected. It was a complex deletion-insertion variant characterized by loss of 17 and gain of four nucleotides. The authors call this variant as "delinsTCCC".

My favorite part of the story is how the authors followed up this variant further to arrive at the causal gene and the variant mechanism. The region containing the delinsTCCC variant turned out to be highly conserved and not that many genes were in the vicinity. The nearest one was ~7kb away, KCNB1, encoding a potassium channel. Studying previously reported GWAS signals at this location, the authors found that there was indeed a GWAS signal nearby, mapped to ECG ST segment changes. The delinsTCCC variant was sitting in between two index variants of the GWAS locus. Analyzing low-resolution chromatin conformation capture data from cardiomyocytes differentiated from induced pluripotent stem cells (iPSCs), the authors found that one of the GWAS variant and delinsTCCC fell within a region that physically interacts with the promoter of KCNB1.

The authors further analyzed epigenetic data based on relevant tissues and cell types from external sources and found that the delinsTCCC did not overlap with any open chromatin region in cardiac cells and the nearest open chromatin region was 12 kbp upstream, sitting between delinsTCCC and the KCNB1 promoter. The authors call this regulatory element as "E-139" (as it was 139 kbp away from the KCNB1 promoter). The authors next studied the expression profile of the E-139 region and found that it is expressed in skeletal muscle and many cancer cell lines, but puzzlingly not in the cardiac muscle. The authors had an intuition that delinsTCCC might have created a new enhancer-like element. As they suspected, a neural network-based sequence prediction model informed that the delinsTCCC sequence indeed have features suggestive of open chromatin, accessible to cardiac-related transcription factors. The authors gene edited iPSC cells to introduce the delinsTCCC variant and differentiated them into cardiomyocytes, and using ATAC seq, they confirmed that the delinsTCCC sequence opens up the chromatin and binds to enhancer-related histone proteins.

The authors went one step further and created transgenic zebrafish lines with the delinsTCCC variant or the nearby E-139 element or corresponding wild type sequence, all placed near a GFP protein with a minimal promoter. As expected, the delinsTCCC model showed fluorescence specifically in the heart cells, confirming that the mutation sequence acts as an cardiac-specific enhancer.

Surprisingly, the E-139 model showed florescence ubiquitously (only mildly in the heart), particularly in the brain. This aligned well with the known neurodevelopmental consequence2 of KCNB1 coding variants. It turns out the KCNB1 potassium channel plays a role in brain development, and probably, has no business in the heart. But the mutation made these potassium channels to express in the heart, resulting in cardiac conduction defects.

Finally, the authors performed chromatin capture assay in gene-edited cardiac cells and tested if the delinsTCCC physically interact with KCNB1 promoter. This is where they found something interesting. It turned out that, unlike E-139 element, the delinsTCCC sequence element doesn't physically interact with the KCNB1 promoter (the physical interaction they initially observed was based on low resolution Hi-C; here they are looking at high-resolution). But interestingly, the physical interaction between E-139 element and KCNB1 promoter became amplified in the presence of delinsTCCC. The findings suggest that delinsTCCC region might not be a conventional enhancer, but might function as a super enhancer that facilitates the interaction between nearby enhancer (E-139) and promoter (KCNB1).

This study is one of the many examples to remind us that we should reconsider our notions about disease-causing variants. Conventional thinking would be to search for variants in regions that are previously annotated as promoter or enhancer in disease-relevant tissues. The study reminds us that disease-variants not only can disrupt already existing regulatory elements but it can also introduce an entirely new regulatory element. Similarly, when prioritizing disease-genes, it is a common practice to narrow down the list by excluding genes that not expressed in the disease-relevant tissues. The study reminds us that disease-variants not only can affect genes that are specifically expressed in disease-relevant tissues but it can also affect genes that are specifically not expressed in disease-relevant tissues.

The authors write "we believe this is the first description of an entirely de novo cryptic enhancer causing a Mendelian disorder", that such type of examples may be more common that one would expect, and with increasing use of whole genome sequencing to diagnose rare diseases, we will learn more of similar noncoding mechanisms in the future.

I ended my previous story with the take home message that the drug discovery field hasn't explored the within-individual biological diversity as much as it has explored the between-individual genetic diversity. And this week, we have another remarkable story that further emphasizes my point.

After completing MD in 2014, I moved from India to Denmark to do a PhD in psychiatric genetics. My interest in psychiatric genetics came from working on a related topic for my MD thesis. That was my entry into the world of human genetics. At that time, I knew little about the field. During my PhD, I learned about the genetics of ADHD, autism, schizophrenia, and other psychiatric diseases. Later, when I began reading broadly, I felt a longing towards certain areas of human genetics like cardiovascular system, endocrinology and metabolism. As I read more, the list grew longer. The latest addition to the list is infectious diseases, particularly the ones that affect children.

Ever since I discovered Jean-Laurent Casanova, an amazing physician scientist at The Rockefeller University, I've been on and off reading about monogenic infectious diseases. Casanova has made tremendous contributions in this area. The first time I learned about him was in early 2022, when I came across two parallel publications in The Journal of Experimental Medicine on the discovery of null mutations in interferon genes that explained the high vaccination-associated deaths in Polynesian Islands and circumpolar regions. I've written about this story previously.

Human immune system has been evolving for over hundreds of thousands of years through continuous arms race against viruses, bacteria, fungi and parasites. For a pathogen, for example, a virus, to infect humans, it has to bypass multiple layers of security measures. It needs to first enter the body, then the cells, then the nucleus if it requires access to host's transcriptional machinery. There are sophisticated sensors in both inside and outside cells that monitor for foreign intruders. Upon detection, they unleash an army of proteins that attack the virus from different angles and chop them into pieces. How scientists first came to know about the genes and pathways that make our extraordinary viral immune defense systems?

I am tempted to say through human genetics. But at least in the case of immunology and infectious diseases that is not entirely true. Even before the reports of gene mutations increasing susceptibility to infections were published, scientists knew a lot about the pathways involved in pathogen immune response through molecular and animal studies. Many of the discoveries of monogenic causes of infectious diseases were in fact driven by prior knowledge of the genes and pathways associated with such diseases. I have highlighted few examples of how mouse genetics played a major role in the discoveries of monogenic causes of systemic lupus erythematosus (SLE).

Autoimmune diseases like SLE represent one end of the double-edged sword that is almost every immune gene. The other end is the immune deficiency. There are countless examples of gene defects that increase susceptibility to viral, bacterial, fungal and parasitic infections. Many of those discoveries were driven by prior mouse genetic findings.

For example, the first reports of genetic defects increasing susceptibility to mycobacterial infection in humans surfaced in 1996. Two parallel studies in NEJM (one of which was from Casanova's group) reported mutations in IFNARG1, encoding interferon gamma receptor 1, abolished immune response to mycobacterial infection, even the ones by attenuated strains. As a result, carriers developed disseminated infections after Bacillus Calmette–Guérin (BCG) vaccination. But even before these reports, scientists knew that deleting any of the genes in the type 2 interferon signaling axis—interferon gamma (IFNG), its receptor (IFNARG1) or its downstream transcription factor protein interferon gamma regulatory factor (IRF1)—makes the mice susceptible to BCG infection. And this knowledge is what motivated Casanova and his team to search for mutations specifically in these genes when studying a child who died of disseminated BCG infection after vaccination.

We examined the five genes coding for interferon-γ, interferon-γR1, IRF1, TNF-α, and TNF-αR1 in a child with fatal idiopathic disseminated BCG infection. We found a mutation of the gene for interferon-γR1. There was no detectable interferon-γR1 on the cells from the affected child. These findings provide further evidence of the importance of interferon-γ in the response to mycobacterial infection.

- Jouanguy et al. NEJM 1996

While human genetic findings in certain scenarios simply confirmed what was previously found through mouse knockout studies, in other scenarios, they brought critical insights that greatly propelled forward the understanding of immune pathways. STAT1 (signal transducer and activator of transcription factor 1) is an important immune protein and is at the cross roads of multiple interferon signaling pathways. Early STAT1 mouse knockout studies focussed mostly on the viral diseases. In 2001, Casanova's group reported the first STAT1 genetic mutation in humans, which increased the susceptibility to mycobacterial infections, but surprisingly, did not increase the risk for viral infections. The two patients with history of disseminated BCG infection were found to have partial STAT1 deficiency due to heterozygous mutations. Just two years later, Casanova's group reported the consequences of complete STAT1 deficiency in humans. Searching for genetic diagnoses in two infants who died of unexplained viral illnesses, the authors discovered that both patients were knockouts for STAT1. The two studies together painted a clear picture of the recessive role of STAT1 in type 1 interferon signaling (induced by alpha and beta), and dominant role in type 2 (induced by gamma).

Our description of infants genetically deprived of STAT-1-mediated responses to IFN-α/β and with lethal viral disease provides the first evidence, to our knowledge, that endogenous IFN-α/β exerts potent antiviral effects in vivo in humans. More generally, our study of humans shows that IFN-α/β is crucial for protective immunity to viruses in natural conditions of infection.

- Dupuis et al. Nat Gen 2003

The search for genetic causes of unexplained infectious diseases has always been guided by prior knowledge of the genes and pathways related to the particular pathogen. And in those searches, often scientists found what they expected to find. But sometimes, human genetics can surprise scientists by revealing something completely new, something extraordinary, opening the door to a new therapeutic mechanism.

In a new paper in Nature, an international team of researchers led by Casanova and Yi-Hao Chan at the Rockefeller University report the discovery of a neuronal membrane protein, tomoregulin-1, that protects the human cortical neurons from herpes simplex viral infection. The gene TMEFF1 encoding tomoregulin-1 was cloned 28 years ago. Until this year, scientists had little idea what this protein does, except that it is specifically expressed in the brain. Now, studying two unfortunate humans who were born without a functional copy of TMEFF1, scientists have discovered that the gene encodes a viral restriction membrane protein that prevents viruses from entering brain cells.

The authors started with a set of 319 individuals with HSV-1 encephalitis (HSE). They searched for mutations in genes previously known to be mutated in HSE, which yielded genetic diagnoses for 30 individuals. The authors suspected that the rest most likely are carrying mutations in genes that are not yet linked to HSE. There is a good rationale for that expectation.

Although primary herpes infections and secondary infections due to reactivation of latent virus are common, herpes virus spreading to the brain during reactivation is rare. It happens once or twice per 100,000 infections per year. And almost every such case is due to some defect in the brain-specific viral immune response pathways. When it comes to viral infections, brain is extraordinarily protected, thanks to evolution. You can appreciate this from the fact that even in conditions like primary immune deficiency caused by defective white blood cells production, the herpes infection do not spread to the neurons. Conversely, in individuals with HSE, rarely the infection spread to other parts of the body. Past genetic findings in HSE has reaffirmed the extraordinary cell-intrinsic nature of antiviral pathways.

Like we have seen in the cases of STAT1 deficiency, genes often has to be completely lost before the virus can infect the brain. Hence, the authors specifically searched for recessive mutations in the participants and found TMEFF1 in which homozygous coding variants were found in two individuals. The first was a French woman who had developed severe HSE at two and half years of age that left behind a large brain lesion visible in brain scan even 16 yrs later. The second was a 19 yr old Turkish man, born to consanguineous parents, who developed severe HSE at 5 years of age. Both recovered from HSE after intravenous acyclovir treatment, though not without neurological sequelae. Exome sequencing revealed that each carried a unique homozygous mutation in TMEFF1—a missense variant in the first and a splice variant in the second.

Given that TMEFF1 is a membrane protein, the authors hypothesized that its function has something to do with the virus entry into the neurons. The authors differentiated wild type and TMEFF1 knockout human pluripotent stem cell (hPSC) into neurons and infected them with HSV-1. As expected, the virus grew rapidly in the knockout cell lines but not in the wild types. The authors observed the same when using patient-derived hPSCs. The HSV-1 virus grew well in patient cells but not in the wild types with intact TMEFF1.

Next, the authors wanted to know if TMEFF1 acted through any of the known HSV-1 related pathways, for example, TLR3 and interferon alpha and beta signaling. It turned out TMEFF1 mechanism is unrelated to any of the known neuron-intrinsic antiviral pathways.

If the TMEFF1 acted at the membrane level, then it should be the first soldier to fight HSV-1 even before any of the others take out their swords. The authors created cell lines defective for known HSV-1 related immune genes like interferon receptor (IFNAR1) and tested if simply over expressing TMEFF1 would protect these susceptible cell lines from HSV-1 infection. It turned out, when over expressed, TMEFF1 confers protection even when key genes like IFNAR1 are inactive. More interestingly, the authors find that you don't even have to express the whole protein, just over expressing part of the protein that protrudes outside the membrane is sufficient to prevent HSV-1 from entering the cells.

HSV-1 like many other pathogens uses receptors to enter the cells. There are five glycoproteins (gB, gC, gD, gH and gL) and three receptor proteins (NECTIN-1, HVEM and PILRa) in the neurons that HSV-1 binds to to enter the cells. So, the TMEFF1 should somehow interfere with HSV-1's interaction with any these receptors to prevent its entry. Through co-immunoprecipitation experiments, the authors found a direct interaction between TMEFF1 and NECTIN-1. The authors further found that it is only the extracellular portion of TMEFF1 that interacts with NECTIN-1. So, the TMEFF1 protects neurons by interfering with HSV-1 binding to its receptor NECTIN1 to enter inside the cells.

What is the therapeutic implication of this discovery? Now we know that TMEFF1 has the ability to restrict HSV-1 entry into the cells, we can create synthetic soluble versions of this protein and use it like a vaccine to prevent HSV1 infection or reactivation. Although the present findings support a therapeutic role for TMEFF1 in HSV1 infection, the authors expect that the approach would be also work in HSV2 infection.

The main take home for me from this study is that biological diversity between cell types in humans offers an unique opportunity to mine targets. In this case, the fact that brain cells have evolved their own antiviral defense mechanisms opens door to a new class of evolution-validated drug designs. The idea aligns closely with our previous story of Big tau expressed only in brain tissues resistant to neurodegeneration in Alzheimer's disease.

I am excited about this emerging new idea of leveraging within individual biological diversity to uncover novel drug targets. I am hoping to see more such breakthroughs in the future.

I love coincidences, particularly the ones related to scientific discoveries. In 1975, particle physicists at the Stanford University and University of California in the US published a discovery of a new elementary particle. Twenty years later, the discovery would fetch Martin Perl, the lead physicist, a Nobel Prize in Physics. Following the ancient tradition of naming things using Greek letters, the newly found particle was named "tau"(τ).

In the same year, scientists from the Princeton University in the US published a discovery of a mystery protein in the brain tissue of pigs. Like how bones form the skeleton of the human body, there are protein complexes--microfilaments, microtubules and intermediate filaments--that form the skeleton of our body cells. Microtubules are long tubular structures built from tiny lego-like pieces called tubulins. There are two types of tubulins: alpha and beta. The two combines to form heterodimers. The heterodimers further assemble into microtubules.

When you crush a microtubules into pieces, you find among the pieces individual tubulin dimers, and also, pieces of ring-like structures each consisting 23 tubulin dimers (which can be further broken down to individual dimers). When you try to polymerize the broken pieces, the larger rings easily come together and transform into microtubules. But the individual dimers don't. Some mysterious factor seems to hold the tubulins together and help them to polymerize. Marc Kirschner and colleagues from Princeton discovered that mystery factor to be a protein. Honoring the ancient tradition of naming things using Greek letters, the newly found protein was named "tau"(τ).

In previous studies of in vitro microtubule assembly, it was shown that depolymerized tubules contain two components: the 6S tubulin dimer, and a 36S species which contains about 23 tubulin 6S subunits arranged mainly as double rings. We observed that only the fractions containing 36S rings would repolymerize to form microtubules; the 6S dimer alone would not. ...

We now report that the ability of 36S rings to assemble into microtubules is due to a salt dissociable factor which resides in the 36S species but is absent in the 6S dimer. We have isolated this factor, a protein which is essential for the assembly of 6S tubulin into 36S rings and microtubules. We propose to call this protein tau (τ) for its ability to induce tubule formation.

—Weingarten et al. PNAS 1975

Both the physics tau and biological tau represents major discoveries in their respective fields. But it was a mere coincidence that they happened to be born in the same year and to be given the same name. If you Google "tau", the first result will be a Wikipedia link to the physics tau. You need to search for "tau protein" to read about the biological tau.

Tau is encoded by MAPT (microtuble associated protein tau) located on chromosome 17. It is highly expressed in neurons where it helps to stabilize the microtubules. Neurons need microtubules not only to maintain their structures but also to communicate with each other. You'd think that tau sounds like a very important protein and without this protein organisms will not survive. Surprisingly, it’s not true. You can delete the MAPT gene from the mice safely without any major consequence, and the mice still will live and breed normally. It was one of the surprising revelations in the Alzheimer's research (Ke et al. Int J Alzheimers Dis 2012.)

Then why does this protein wreak havoc in the brains of individuals with conditions such as Alzheimer's disease, frontotemporal dementia, progressive supranuclear palsy (PSP) and Pick's disease, etc.? Tau proteins are complex proteins with many structural domains that make them inherently unstable. Many factors can destabilize tau and make them stick together to form tau tangles. Mutations that result in abnormal secondary and tertiary structures (misfolded tau proteins) can lead to sticky tau. Certain splicing mutations add or remove certain parts of the tau protein and consequently, increase and decrease the sticky tendency. The tau proteins have certain amino acids like serine and threonine in their structure that can accept a phosphate group. Microtubules are dynamic structures. Correctly adding and removing phosphate groups to and from the tau protein, respectively, is essential for the tau to assemble and dismantle the microtubules. Too many phosphates can kick the tau off the microtubules, resulting in tau aggregation. A worst thing about tau aggregation is it can propagate. As one tau protein start sticking to an another, the next and next tau proteins follow suit and the chaos can gradually spread across the brain like it occurs in Prion diseases.

One interesting feature of the MAPT gene is that its nascent mRNA transcript are spliced in different ways to producing a variety of isoforms that subtly differ from each other. In the context of central nervous system, splicing of three three exons (2, 3 and 10) determines the isoforms. Based on presence or absence of exons 2 and 3, the protein is labelled 0N (neither exons are included), 1N (only exon 2 included) and 2N (both exons 2 and 3 are included). The exon 10 codes for a repeat region that forms the microtubule-binding interface. Based on the presence or absence of exon 10, the protein is labelled as 4R (four repeats) or 3R (three repeats)1. So, alternative splicing of the three exons (2, 3 and 10) gives rise to six different isoforms: 0N3R, 0N4R, 1N3R, 1N4R, 2N3R, 2N4R.

While the six isoforms are mainly found in the central nervous system (CNS), in early 1980s, scientists came across a new high molecular form of tau protein in the peripheral nervous system (PNS) tissues like spinal cord, spinal ganglia, peripheral nerves etc. This larger isoform was named "big tau" or "PNS tau". The exon contributing to the extra length of big tau was cloned in the early 1990s (Goedert et al., Couchie et al. PNAS 1992). As the MAPT exons were already numbered by then, the newly found exon located between exons 4 and 5 was given the number 4A (shown as black rectangle in the figure below).

For some reason, the big tau never attracted much attention from neuroscientists for more than 25 years after its discovery (Fisher eNeuro 2023). Now the big tau is making a comeback. A new study from the lab of Huda Zoghbi, a world-renowned neuroscientist known for her pioneering work in conditions like Rett syndrome, spinocerebellar ataxia 1 etc., reveals the extraordinary qualities of big tau, nominating it as a potential therapeutic target for tauopathies.

In the preprint, Chloe Chung et al. reports that, on contrary to the popular belief, the big tau is expressed in the brain. The authors show this both in mice and humans. It turns out the brain regions expressing the big tau are the ones that are typically spared from neurodegeneration in the Alzheimer's disease. Using western blot, the authors show that big tau is abundantly expressed in the mouse cerebellum and brain stem. The authors further show the same in the postmortem brains of Alzheimer's patients and controls. Specifically, the authors show the big tau is more abundant in the cerebellum of Alzheimer's brain than that of controls.

Motivated by the abundant expression of big tau in brain and its unique structure with an N terminal end that is three times longer than the regular tau isoforms, the authors studied the biochemical properties of big tau. As I mentioned before, one of the key chemical properties of tau is phosphorylation. The level of tau phosphorylation increases with age, and it also increases in the diseased state. By measuring the phosphorylation levels of big tau in younger and older mice, the authors show that big tau is less subject to age-associated phosphorylation.

One of the ways through which the neurons clear the tangled tau proteins is ubiquitination, a process of adding little protein tags to signal the cellular protein clearance system. The ubiquitin are added to lysine residues, and it turns out the big tau contains 20 additional lysine residues than the regular tau. The authors show that the big tau gets more easily ubiquitinated and consequently, gets efficiently cleared, compared to the regular tau. And this property makes lifespan of big tau much shorter than the regular tau.

One of the main problems with tau proteins in Alzheimer's disease is they easily fall off the microtubules. When you centrifuge the microtubules at high speed, the tau that are not tightly bound to microtubule get separated, which can be collected as a supernatant. The authors show that, unlike regular tau, the big tau doesn't get separated from microtubules. After ultra-centrifugation, they are still found bound to microtubules at the bottom. What makes the big tau bind more tightly with microtubules? Using a combination of bioinformatics and microtubule binding assay, the authors show that big tau possess extra microtubule binding motifs. The binding affinity is so strong, in a competitive binding assay, the big tau outcompetes regular tau even when present in small amounts.

Finally, the authors test the most critical property of the tau proteins—aggregation tendency. There are certain mutations that increase the aggregation tendency of tau proteins. The authors show that even if you introduce such a mutation, the big tau refuses to aggregate, both in vitro and in vivo (in the mouse brain).

Overall, the authors demonstrate that big tau has many properties that make it resistant to Alzheimer's disease pathology: it resists hyperphosphorylation, it gets easily cleared, it binds tightly to microtubules and it refuses to aggregate, even in the presence of aggregation-promoting mutation. How amazing is that? I am surprised that this protein has evaded scientists' attention for nearly 3 decades since its discovery.

The discovery has a major therapeutic implication. Tau is an important drug target of interest for many companies. Many modalities are being tested. Many are developing antibodies to bind and clear the tau tangles. Some are developing small molecules to prevent tau aggregation. Some are developing siRNAs to knock down MAPT expression itself. The new findings opens a new therapeutic idea. If big tau is resistant to aggregation, then one can try a non-destructive way of treating tauopathies, for example, instead of completely shutting off the tau production, one can simply nudge the neurons using splice modulators to produce big tau instead of regular ones. Its not as easy as it sounds, but definitely worth trying. I am sure many companies will do.

One thing that came to my mind when reading this paper was the recent Casgevy approval for the treatment of sickle cell disease2. Casgevy is a CRISPR-based treatment. The patients' blood stem cells are taken out of the body, gene edited to reduce the expression of a gene called BCL11A and then put back into the patients. The BCL11A is a key regulator of fetal to adult hemoglobin switching. By reducing the BCL11A expression, Casgevy increases the fetal hemoglobin production, protecting the red blood cells from sickling and thereby, increasing the quality of life of the patients. I wonder if there exists any splicing regulatory protein that can be targeted to increase the big tau production in the neurons.

One key take away from this work for me is, we human geneticists always go in search for genetic diversity across humans to find drug targets. But there is incredible genetic diversity within individual human brains that has been largely unexplored to date. The current work from Huda's lab and the recent work on Huntington's disease from Steve McCarroll's lab are great examples. I expect we will see more such work in the near future.

One of the biggest mysteries that scientists were wrestling with during the early days was the puzzling disconnect between size of the organisms and their genomes. Larger animals have smaller genomes, whereas some plants, insects and even some unicellular organisms (e.g. Amoeba dubia) have humungous genomes. Just a couple of months ago, scientists found on the forest floors of an Australian island a plant species with the largest genome ever known to humans (Fernández et al. iScience 2024). Tmesipteris oblanceolata, a tiny fern species endemic to eastern Australia has a genome that is 50 times larger than that of ours (NYT article).

The puzzling relationship between the genome size and organismal complexity, described as 'C-value paradox', is now explained by the fact that a major part of genomes are 'graveyards' of repetitive elements containing fossilized genomes of ancient viruses. More than 45% of the human genome are repeat elements filled with transposons, aka, jumping genes. Transposons jump from one part of the genome to another and, in the process, edit, delete or create new genes, bestowing upon its host new phenotypes. They are one of the major drivers of human evolution.

Just a month ago, there was a big breakthrough in genome editing where researchers from the Arc Institute in the US and University of Tokyo in Japan uncovered the molecular mechanism through which an evolutionarily old, most simplest forms of mobile elements called insertion sequences (IS) found in bacterial and archaeal species jump across the genome(s) (Durrant et al. Nature 2024, Hiraizumi et al. Nature 2024). The breakthrough part is the realization that a component of this ancient transposition machinery was a noncoding RNA that bridges the transposon itself (bound to its one loop) with the target DNA integration site (bound to its other loop). Both the loops of this "bridge RNA" are programmable. So, by altering the RNA loop sequences, one can trick the recombinase enzyme that is part of the machinery to make large structural changes at desired target sites in the genome. You can basically cut out a large piece of DNA, or remove a piece, flip it and put it back, or even insert a whole new gene into the genome.

Okay, let's not drift into the gene editing literature. Coming back to our topic, that is, human genetics, in a recent work published in Nature Genetics, Po-Ru Loh, Steve McCarroll and team at the Broad Institute in the US report how a transposon sitting within a skin pigmentation gene ASIP has driven the evolution of skin color in humans and their ancestors.

ASIP encodes 'agouti signaling protein', a paracrine hormone produced by the skin cells. The agouti protein competes with melanocyte stimulating hormone (MSH) to bind to melanocortin 1 receptor (MC1R) in the melanin-producing skin cells. Agouti hormone blocks the activation of MC1R by MSH in the skin and slows down the production of eumelanin (black-brown pigment), skewing the balance towards pheomelanin (yellow-red pigment). Agouti protein has a fascinating literature spanning multiple species including mice and humans, worthy of a whole book. The first time I read in detail about this protein was last year when I came across a paper that reported the discovery of a structural variant in ASIP as a rare monogenic cause of morbid obesity in humans. I've told this story in a previous post. It's a fascinating one; do check it out, if you haven't already.

As was the case with many of the metabolic genes such as leptin (discovered from ob/ob mice), our knowledge of agouti gene and its function came from mouse genetics. Spontaneous mutations in agouti gene cause fur color variations in mice. This knowledge led to the successful cloning of the human agouti gene in 1994 (Kwon et al. PNAS 1994). Although candidate gene studies have reported association between genetic variations near ASIP and skin pigmentation (Kanetsky AJHG 2002), the first robust evidence came from a genome-wide association study (GWAS) by deCODE genetics in 2008 (Sulem et al. Nat Gen 2008). Studying around 5000 Icelanders, Sulem et al. identified a variant near ASIP that was strongly associated with skin sensitivity to sun, freckling and red hair, a set of phenotypes reminiscent of the genetic associations previously reported for MC1R. In the same issue, two parallel papers reported the association of ASIP locus with skin cancer risk. Although, the causal gene was pretty much clear, the none of the research teams were able to isolate the causal variant. Subsequent studies based on European population replicated the ASIP GWAS signal and also found that the risk variants increase the ASIP gene expression in the skin. However, the causal variant at this locus have remained unknown. Interestingly, later GWAS of skin color in African and East Asian ancestries failed to identify a signal near ASIP, hinting that the locus is specific to European ancestries. In line with that, a study in 2016 found the ASIP locus as one of the regions of the human genome that underwent positive selection in Europeans (Field, Boyle, Telis, et al. Science 2016).

Po-Ru Loh and his team are experts in building imputation models based on common variants and using the imputation models to predict rare and complex genetic variations like rare coding variants and structural variants that otherwise can be identified only using sequencing, some only using long-read sequencing. They have been digging out fascinating variable number tandem repeats (VNTRs) in the coding and noncoding regions that were driving some of well-known GWAS associations, e.g., ACAN locus associated with height, TMCO1 locus associated with glaucoma, EIF3H locus associated with colon cancer (Mukamel, Handsaker, et al. Science 2021 & Cell 2023). Adding to this list now is the ASIP locus associated with skin color and skin cancer.

A characteristic shared by all GWAS loci previously found to be driven by structural variants was their large effect size. Given the large effect size associated with a common variant, Kamitaki et al. suspected that the ASIP locus might be hiding a structural variant underneath. So, they started digging. First they searched through the long-read sequencing data of 32 individuals made publicly availably by the Human Genome Structural Variation Consortium. One of the 32 genomes that contained the ASIP risk allele also contained a large structural variant in the same chromosome: a 3.3 kilo basepair insertion that was absent in the rest of the 31 genomes. On a closer look, the authors realized that this extra piece of DNA contained a mobile element called SINE-VNTR-Alu (SVA). SVAs are evolutionarily young and seen only in primates, including humans. They are retrotransposons, that is, they jump across the genome not through "cut and paste" approach, but through "copy and paste" approach using an RNA intermediate. So, they leave behind a copy of their own every time they jump. Further digging into the genome, the authors found that this transposon was not alone. It had a clone of its own a little further down but buried in the opposite direction, like a mirror image. Unlike the former SVA (the authors call this 'SVA F1'), which is polymorphic, meaning it is present in some humans and absent in others, the latter (SVA F) is monomorphic, meaning, it is present in all the humans. Here, things got more interesting as it appeared that not just one but two transposition events had occurred at this locus during human evolution.

To understand the functional consequence of these mobile insertions, the authors analyzed the RNA sequencing data from the GTEx database, comparing the gene expression across individuals with and without the SVA F1. Skin tissue of individuals heterozygous for SVA F1 had more ASIP expression compared to non-carriers. On closer inspection, the authors spotted one particular ASIP transcript that was decreased in expression in SVA F1 carriers as opposed to rest of the transcripts that were increased in the carriers. This aberrantly spliced transcript was almost absent in individuals homozygous for SVA F1. It turned out this aberrant splicing was caused by a splice acceptor site introduced by the non-polymorphic SVA F (SVA F1's buddy).

Long story short, the story unfolds like this: sometime during the evolution after the human split from the great apes but before they split from Neanderthals (~500,000 years ago), the SVA F retrotransposon jumped inside the intron of ASIP, introducing an aberrant splicing and reducing the ASIP expression. The resulting darkening of skin gave humans survival advantage, as the humans have already started losing their body hair and getting exposed to a lot of sun light. The evolutionary advantage was so strong, only the humans who carried the SVA F insertion survived, and the allele became fixed in the lineage. Then, sometime very recently during the evolution after the great migration out of Africa, a second transposition event occurred; the SVA F1 jumped straight into the same intron in the inverse direction, somehow1, releasing the ASIP expression block caused by SVA F. The resulting lightening of skin too gave some evolutionary advantage to the humans who settled over the northwestern regions of Europe. The selection pressure was strong that the SVA F1 allele rapidly rose to high frequency (7-8%) in the northwest European populations; the allele is almost absent in Asians and African populations. As always, the benefit came at a cost—increased susceptibility to skin cancer. Carrying one allele of SVA F1 increases the risk of skin cancer by 1.2 fold and carrying two allele by 1.4 fold.

Overall, it's an another remarkable work by Po-Ru Loh, Steve McCarroll and team from Broad Institute. If you think about it, the association of the structural variant itself wasn't that difficult to find. The structural variant was part of the human genome annotation. If one has looked at the genetic associations of all known structural variants with skin color or skin cancer risk, they would have found it. Yet, no one did all these years.

As the authors put it, the human genomics field has ignored structural variants for a long time. There are likely a whole lot of similar evolutionary stories buried within the noncoding genome waiting to be told. Let's hope we will hear more of them in the near future.

The human genome has around 18,000 to 20,000 genes that code for proteins. I often wonder how much the field has learned about the function(s) of each of these genes. One thing about human genetics that has never ceased to amaze me is how naturally occurring spelling errors in the gene sequences in the population shed light on the normal functions of the genes. The value of something is appreciated only when it is lost. Many of the genes in the human genome were known to exist today only because scientists encountered humans in the past in whom such genes didn't exist, functionally speaking.

One class of genes that I find particularly interesting are the ones that encode transporters and ion channels. Millions of molecules are on a constant move in the human body, and often their movements are facilitated by transporters and channel proteins. They let molecules shuttle in and out of cells and cell organelles. Sometimes the cells merely act as passages through which molecules enter into or exist out of the human body to conduct their day-to-day businesses. Genetic mutations that partially or completely block such passages create traffic jams, leading to crowding of the passengers either inside or outside the body and send signals to scientists, setting the stage for new discoveries.

One channel protein that probably comes to everyone's mind with the mention of human genetics is cystic fibrosis gene CFTR. It encodes chloride channels in the cells lining the walls of the ducts that allow air to flow into the lungs, digestive enzymes to flow through the pancreas and sweat to be released from sweat glands. Loss of this channel blocks the movement of chloride and sodium and consequently, water, across the ductal epithelial cells, dehydrating and driving up the viscosity of the ductal contents.

One of the key pieces of the cystic fibrosis puzzle that helped scientists triangulate the genetic cause appeared during a hot summer in 1948 when a heat wave swept through the New York. Paul di Sant'Agnese, a pediatrician at the New York-Presbyterian Hospital, noticed that cystic fibrosis patients were disproportionately exhausted by the heat than other patients in the hospital. This observation led to the discovery of excessive loss of salt (sodium and chloride) in the sweat in cystic fibrosis patients, and subsequent realization that the core problem in cystic fibrosis might be related to ion channels.

The fact that the sweat glands do not retain salt effectively in hot weather, during salt restriction or when DCA [deoxycorticosterone acetate] is administered, points to a disturbance in the sweat glands themselves. This abnormality may not be pathognomonic of cystic fibrosis of the pancreas, but it is an expected finding.

These disturbances in sweat composition together with preliminary observations of abnormal salivary gland secretion suggest that the secretory activity of many and perhaps all exocrine glands is affected in cystic fibrosis of the pancreas. The term "mucoviscidosis" does not seem justified in view of the fact that at least two nonmucus-secreting structures are affected in this condition

Sant'Agnese et al. Paediatrics 1953

One of the examples that first come to mind when I think of human genetics and transporters is the discovery of uric acid transporters. Uric acid is an end product of purine metabolism (breakdown of adenine and guanine nucleotide bases). Uric acid is primarily excreted through kidneys. The blood levels of uric acid are determined by the balance between its rates of production and excretion. Abnormally high levels of uric acid in the blood is bad. It leads to deposition of uric acid crystals in the tissues and joints, resulting in gouty arthritis.

The knowledge of gout itself is remarkably ancient, known to exist for thousands of years. The link between uric acid and gout is also known for nearly two centuries now. However, the knowledge of key transporters in the kidneys that titrate the uric acid levels in the blood remained a mystery until the early 2000s. Then the human genome project happened, laying the map of human genome for scientists to navigate through in pursuit of new genes.

The uric acid that escape through the glomerulus of the nephrons in the kidney gets reabsorbed back into the system through the cells lining the tubular portions of the nephrons. This process is accomplished by two transporters: one embedded on the side facing the inside of the tubes (apical surface) and the other facing the blood circulation (basolateral surface).

The transporter on the apical surface, SLC22A12, was discovered in 2002 through a systematic search across the genome for transporter-like gene sequences, followed by molecular characterization and animal experiments (Enomoto et al. Nature 2002). Only after learning the apical surface location of SLC22A12, scientists realized that there should be another similar transporter on the basolateral side that shuttles the uric acid into the blood circulation (shown as a question mark in the illustration below).

In 2008, three genome-wide association studies (GWAS) in parallel reported the discovery of a strong genetic association between SLC2A9, encoding a presumed glucose transporter GLUT9, and serum uric acid levels (Wallace et al. AJHG 2008; Vitart et al. Nat Gen 2008; Döring et al. Nat Gen 2008). It turned out, GLUT9 was primarily a urate transporter, the mystery protein sitting on the basolateral side of the renal tubules and working in partnership with SLC22A12 to reabsorb uric acid from the glomerular filtrate.

Today, we know there are a bunch of urate transporters sitting on the renal tubular cells and genetic variations near genes encoding almost all of these proteins have a striking influence on the uric acid levels in the blood in humans, which you can beautifully see as skyscrapers in a GWAS Manhattan plot. Even if molecular biologists did not discover these transporters early on, they would have eventually come to light through GWAS.

There are hundreds of metabolites like uric acid in the blood, and have we figured out all the transporter genes shuttling these metabolites across various sites in the human body? A recent discovery reminds us the answer to that question.

See, discoveries of the chloride channel and urate transporters were heavily inspired by the associated diseases. The discoveries happened as scientists were aggressively studying the root causes of the cystic fibrosis and gout. There are probably hundreds of transporters whose roles have not yet been realized because disruption of their functions either do not lead to major disease states or the associated diseases have not yet been linked to the causal metabolites. And all it takes is a small hint, for example, as in the recent discovery where scientists picked up on a genetic signal linked to plasma and urine choline levels and ended up discovering the human mitochondrial choline transporter.

In the past few months, at least three research groups have independently discovered that there exists a transporter protein, encoded by SLC25A48, in the mitochondrial membrane that is required for choline to get inside the mitochondria to do its business (Patil et al. medRxiv 2024; Khan et al. Nat Gen 2008; Verkerke et al. bioRxiv 2024). When this transporter doesn't work, choline levels spike up in the blood and urine.

Choline is an essential nutrient, a vitamin-like chemical required for many important biological processes such as synthesis of phosphatidylcholine (a component of cell membrane), synthesis of neurotransmitters (e.g. acetylcholine) and one carbon metabolism (e.g. folate metabolism).

The fascinating part of the discovery of choline transporter is that scientists weren’t even looking for them. Probably, no one even realized that there existed another transporter for choline (a few choline transporters were already known to exist). That's the beauty of hypothesis-free studies such as GWAS. You go in there blind and stumble upon something new, pointing to an important biology.

One of the three research groups that discovered the choline transporter is from the University of Freiburg in Germany, who conducted GWASs of 1,916 plasma and urine metabolites in 2023 (Schlosser et al. Nat Gen 2023). One of the 1,299 genetic signals they identified was found near an orphan transporter protein SLC25A48, associated with urine choline levels. They followed up this finding further and found that the genetic variants near SLC25A48 are not only mQTLs for choline in the urine, they are also mQTLs for choline in the plasma and eQTLs for SCL25A48 in the brain tissue. They decided to dive deep into the signal and performed various functional experiments and arrived at the conclusion that SLC25A48 is a choline transporter located in the inner mitochondrial membrane, and a major regulator of plasma choline levels.

The second group is from the Rockefeller and Vanderbilt Universities in the US, who also spotted the SLC25A48 signal through GWAS of plasma metabolome. The authors first grouped the significant associations into two groups: ones for which the causal genes are clear based on prior literature and the ones for which causal genes aren't clear, hence provide opportunities to uncover new biology. They decided to study the top most signal in the second pile: SLC25A48 associated with plasma choline. Then followed a series of in vitro experiments, which led to the revelation that SLC25A48 is a mitochondrial choline transporter. The authors also report possible health consequences of SLC25A48 dysfunction by performing a phenome-wide association study of rare loss of function variants and common variants near SLC25A48. The associations include diseases of nervous system, bone, tooth, skin and blood vessels.

The third group is from Beth Israel Deaconess Medical Center and Harvard Medical School in the US, who landed on SLC25A48 without the help from human genetics. The authors performed a proteomics analysis of mouse brown adipose tissue in the context of diet-induced obesity and ended up discovering SLC25A48, which was significantly upregulated in the brown adipose tissue in a high-fat diet. Long story short, the authors studied the mitochondrial respiratory function in Slc25a48 knock out mice and somehow discovered the link to choline and arrived at the same conclusion as other two groups.

It is fascinating to see a biological finding remains hidden for a long time and then suddenly it comes to light with a bang. Studies like these are reminders that there is still a lot more exciting biology remain unknown to the world, hidden beneath massive amounts of genomic data, waiting to be discovered.

Until I was in academia, I worked within a small group of researchers who all pretty much had the same expertise as I had and so, I was never used to explaining certain genetic concepts during the scientific discussions. This worsened my—what advocates of good writing like Steven Pinker describe as—"curse of knowledge". I only realized this after I started working in industry. In my day job, I communicate genetic results to scientists whose expertise differ drastically from mine. One of the important concepts that I often find myself explaining again and again to my neuroscientist colleagues is “mutational constraint”, that is, why certain genes are depleted of deleterious mutations. We often get requests to check if mutations in a gene is associated with any disease. Being in a neuroscience area, often such genes will be related to brain development and so, will be constrained. As a result, we won't find that many carriers of deleterious mutations, despite having access to a genotype-phenotype database of more than 2 million individuals. Every time, when we deliver the gene look up results, we go through certain basic statistics such as expected vs observed loss of function variants and conclude that the gene is mutationally constrained and so, we couldn't find an enough number of carriers to confidently interpret the phenotypic associations.

Genetics of human brain development has always fascinated me. It is one of those areas of human genetics where you get a first row seat to watch natural selection in action. When I write about papers on Twitter, I never miss an opportunity to share this famous world warplane illustration, while explaining results related to mutational constraint.

One of the uses of sequencing the general population in large numbers is to predict which parts of the genome are critical for human survival. The logic is that if you see fewer variations in a part of the genome than what you'd expect to see in a sample of, say, 500,000 humans, you infer that that part of the genome is important for survival. Individuals who were born in the past with mutations within that part of the genome had probably not survived. The inspiration behind this logic comes from the famous World War II story of Abraham Wald. Here is how a 2018 paper on mutational constraint begins:

During World War II, Abraham Wald and the Statistical Research Group optimized the placement of scarce metal reinforcements on Allied planes based on the patterns of bullet holes observed over many sorties. Wald famously invoked the principles of survival bias to infer that armor should be placed where bullet damage was unobserved, since the observed damage came solely from planes that returned from their missions. Wald reasoned that planes that had been shot down likely took on critical damage in such locations.

Employing similar logic, we sought to identify localized, highly constrained coding regions (CCRs) in the human genome. . .

Havrilla et al. Nat Gen 2018

The map of mutational constraint across the genome is critical for interpreting pathogenic variants in patients with neurodevelopmental disorders. Often, clinical geneticists encounter mutations in their patients that were never seen before in any humans sequenced to date. Not even their parents. Such mutations are called private mutations, and if their location map to constrained genomic regions, then they are highly likely to be deleterious in nature. They have likely arisen de novo as a result of some mutational event, either in the sperm from their father or the egg from their mother at some point before fertilization.

A newly born human brings to this world, on average, 70-90 new mutations. The human genome is approximately 3 billion base pairs long, 6 billion when you consider both the pairs of chromosomes. De novo mutations occur at random. Well, not entirely random. I'll come to that in a moment. But for now, let's assume that it is entirely random. So, theoretically, the odds of a particular base pair to get mutated in a random mutation event is 1 in 6 billion. Empirical estimates suggest that each human experience 70 to 90 random mutation events. Using probability calculations, the odds of a particular base pair to get mutated by one of the 70 to 90 mutation events would be 1 in ~85 million to 1 in ~66 million, which is pretty close to the empirical estimate of 1 in ~83 million (1.2e-8). So, here comes the interesting part. The odds of two individuals to carry a private mutation in the exact same position would be 1.2e-8 x 1.2e-8, which is approximately 1 in 7 quadrillion chance. In other words, theoretically, it is nearly impossible for such events to occur. But what we have learned from sequencing humans in large numbers is that is not entirely true. There are regions in the genome that beat the odds of 1 in 7 quadrillion and display the exact same de novo mutation in two unrelated individuals from different parts of the world.

When the first release of 150,000 UK Biobank whole genome sequences was analyzed by deCODE genetics, the researchers found that around 27.7% of the ~200,000 de novo mutations identified previously in the Icelandic population were also seen in UK Biobank participants. This is not surprising as we know that certain regions of the genome are prone to mutate than other regions due to reasons such as chromatic structure, sequence context etc. But still, chances of more than one individual carrying an identical de novo mutation is extraordinarily rare.

. . . These recurrence phenomena have been described in other sample sets using sharing of rare variants between different subsets. We used a de novo mutation set from 2,976 trios in Iceland to assess recurrence directly, as variants present in both that set and the UKB must be derived from at least two mutational events. Out of the 194,687 Icelandic de novo mutations, we found 53,859 (27.7%) in the UKB set, providing a direct observation of sequence variants derived from at least two mutational events. As expected, we found that CpG>TpG mutations are the most enriched mutation class in the overlap, owing to their high mutation rate and saturation in the UKB set.

Halldorsson et al. Nature 2022

Okay, now let's calculate the odds of a same de novo mutation to occur in three individuals. That would be 1 in 578 sextillion chance (5.7e-23). Now, let’s calculate the odds of a same de novo mutation to occur in 89 individuals, which would be 1.6e-705, in other words, astronomical. Well, researchers have recently identified a spot in the genome that beats this astronomical odds.

Analyzing the whole genomes of more than 10,000 individuals with undiagnosed neurodevelopmental disorders, an international team of researchers led by Nicky Whiffin from University of Oxford identified a de novo single base pair insertion variant (n.64_65insT) in 89 individuals. On further analysis, the authors realized that the insertion mutation sits within an 18 base pair sequence of a noncoding RNA gene that is highly depleted of variants compared to its surrounding region in the healthy population. On the other hand, the same region is enriched for de novo mutations, particularly insertion mutations, in individuals with NDDs. In total, the authors identified 119 NDD patients with a de novo mutation within the 18 base pair region. These estimates suggest that around 0.4% of all NDDs across the world (which would correspond to many thousands of NDD patients) could be explained by de novo mutations in this 18 base pair region, which is a mere 0.0000006% fraction of the full human genome. This makes RNU4-2-related NDD as a separate disease entity of its own. It will be a major disease candidate for drug development by many biotech companies in the next decade.

If you think about it, there are two opposite forces in action at this 18 base pair region within the RNU4-2 gene. On one side, natural selection at the population level is removing the variant from the gene pool as the carriers likely die early in life and even if they lived beyond the reproductive age, they rarely procreate. On the other side, some other force keeps mutating this spot in humans again and again, putting back the carriers that natural selection has been actively weeding out. It’s a force that is beating the astronomical odds of observing a same de novo mutation in thousands of humans around the world. The question that next needs to be answered is what that mysterious biological force is? There is already a hint.

Among the RNU4-2 mutation carriers, the authors were able to trace in 54 individuals the parent of origin of the de novo mutations, that is, if the mutation event occurred in the sperm or in the ovum. The authors were surprised to find out that in all 54 individuals, the RNU4-2 de novo mutations were maternally derived. This, again, astonishingly beats the expectation of 80% paternal origin. There lies a some fascinating biology behind why female eggs are prone to hyper mutate at this genomic region. Perhaps, loss of RNU4-2 role in gene splicing favors survival of eggs? We will find out soon.

PSEN1 encodes presenelin 1, a component of a large multiprotein enzyme complex called gamma-secretase. This enzyme complex is what cleaves the amyloid precursor protein (APP) resulting in amyloid beta peptides that aggregate to form the amyloid plaques, seen in the brains of AD patients. Certain missense mutations in PSEN1 cause a severe form of autosomal dominant AD with very early onset, discovered in the mid 1990s. The largest of AD families affected by PSEN1 mutations resides in Antioquia in Colombia. It's a large pedigree of nearly 6000 members of which around 1000 individuals carry a pathogenic missense mutation, E280A1, in PSEN1. The Antioquia Alzhiemer's family has attracted a great deal of interest from AD researchers. Nowhere else in the world one can find so many AD patients, all caused by a single mutation. So, naturally, scientists have been closely watching this kindred to identify any outliers. In 2019, researchers from Harvard Medical School and Universidad de Antioquia (Arboleda-Velasquez, Lopera, O’Hare, et al. Nat Med) reported one such outlier: a Colombian woman with E280A mutation who was destined to develop Alzheimer’s in her late 40s, but never did until her 70s. It then turned out this woman was gifted with a rare APOE mutation2 called Christchurch mutation in the homozygous state. The Christchurch mutation reduces the APOE function3. It was the first ever genetic evidence to support the concept of targeting APOE protein to treat Alzheimer's disease in the absence of APOE4 mutation.

Although APOE has always been a therapeutic target of interest, the discovery of the Christchurch mutation dampening the PSEN1 mutation effect hinted that the target population for an APOE-targeted medicine could involve more than APOE4 carriers. However, that was just an n=1 study. The authors wanted to know if this protective effect is seen even in heterozygous carriers of Christchurch mutation. So, they screened the Antioquia pedigree and found 27 carriers. They studied retrospectively the longitudinal clinical data of these individuals, including the autopsy reports of four deceased individuals. The authors (Quiroz, Aguirre-Acevedo, Vasquez, et al. NEJM) have reported their observations in NEJM recently. The bottom-line is that E280A carriers who also carried one copy of Christchurch mutation appear to have a relatively less severe Alzheimer's disease compared to other E280A carriers. The age of onset of cognitive impairment was delayed by around 5 years. Brain imaging and autopsy examination revealed limited tau pathology4 and neurodegeneration.

One of the interesting findings in both the earlier and current reports is the presence of high amyloid beta plaques in the brain despite a limited tau pathology. The postmortem brain examinations of both the homozygous woman in the earlier report and four deceased heterozygous carriers in the current report show the brains of these individuals had high amyloid beta plaques comparable to that of other E280A mutation carriers. So, the protective effect is seen despite high amyloid beta burden, which hints that APOE targeting might work even after the patients start making high amyloid beta plaques in their brain. The finding also adds to the long-running debates on amyloid hypothesis in AD, perhaps hinting that the AD treatment approaches shouldn't be exclusively focussed on clearing amyloid deposits in the brain.

The APOE mutation modifying the penetrance of PSEN1 mutation is an example of trans-genetic modifier. That is, the modifier (APOE) and modifiee (PSEN1) genes are located on different chromosomes. Recently, there have been also reports of cis-genetic modifiers in AD. In the last year round up, I highlighted a preprint by Michael Greicius and colleagues from Stanford University on APOE loss of function mutations eliminating the pathogenic APOE4 alleles. The paper was published in Neuron a few months ago. It's a nice example of cis-genetic modifier.

APOE4 is the most common genetic risk factor of AD, with E4 homozygosity is now widely recognized as a highly penetrant, monogenic cause of AD. One of the therapeutic approaches drug developers are trying is to treat AD (caused by APOE4 mutation) by inhibiting APOE. Hence, there is interest in studying the effect of APOE loss of function mutations in the population. However, individuals with loss of function mutations in APOE are rare. The Stanford researchers went one step further and looked for an even rarer set of individuals: those with both APOE loss of function mutation and APOE4 mutation in the same chromosome. Surprisingly, they did manage to find a handful of such individuals by scanning through an AD cohort. The authors found that individuals who carried APOE loss of function mutation in the same chromosome as E4 allele did not develop Alzheimer's until their old age and did not show evidence of Alzheimer's pathology in the brain images or in the CSF. The major limitation of this paper is there were only a handful of carriers, and none of them were E4 homozygotes. Since E4 heterozygosity is not that highly penetrant as E4 homozygosity, it's critical to demonstrate protective effect in homozygotes. Given now there is a consensus that E4 homozygosity is near 100% penetrant, identifying even one individual with E4 homozygous genotype and a APOE4 loss of function variant and showing that the AD severity is reduced in this individual will strengthen the case for APOE-targeted therapeutics for AD.

Genetic modifier is an emerging topic in the drug development field lately. In fact, one of the main goals of drug companies investing in human genetics is to identify genetic modifiers that can validate their drug designs. With the rapidly growing genetic databases, interesting examples are beginning to emerge. I’ve discussed one such example in my earlier posts: a loss of function missense variant, N264K, in APOL1, abolishing the effect of kidney disease-causing APOL1 mutation5, which validated the therapeutic approach of inhibiting APOL1 to treat chronic kidney disease.

Another disease area where genetic modifiers are a hot topic is repeat mutations-associated neurodegeneration like Huntington’s disease, cerebellar ataxia, amyotrophic lateral sclerosis (ALS) etc. Our ability to investigate the genetic modifiers of repeat mutations in the population so far has been limited due to unavailability of whole genome sequencing datasets, as often the disease-causing repeat mutations are located in the noncoding genome. As a result, search for modifier variants focussed heavily on common variants. For example, a genome-wide association study of age of onset of Huntington’s disease has reported a common variant near MSH3 that encodes a DNA mismatch repair protein. The finding aligns with the emerging interest on targeting mismatch repair genes to prevent somatic repeat expansion of Huntington’s mutation. Large-scale sequencing projects focussed on Huntington’s disease are currently in progress. Hopefully, such efforts will reveal more genetic modifiers of therapeutic value.

One of the limitation of studying genetic modifiers, particularly in the disease areas of neurodegeneration, is lack of good endophenotypes. In the case of APOL1, the researchers have a great surrogate marker which is in vitro trypanolytic activity. Another example where availability of a good enophenotype accelerated drug target discovery is sickle cell disease (SCD). The BCL11A was discovered through studying the genetics of blood fetal hemoglobin fraction, which is a strong causal predictor of disease progression in SCD. Remember, SCD is a monogenic condition caused by a single point mutation in the HBB gene. So, the common variants near BCL11A discovered via GWAS are in fact trans-genetic modifiers of HBB mutation. If it were the case that the researches had to discover genetic modifiers of HBB mutation only by studying by correlating genetics with clinical markers of disease progression of SCD, we wouldn’t have got an FDA-approved CRISPR-medicine for SCD today. Speaking of which, scientists are already in track to developing small molecule drugs that can up regulate fetal hemoglobin in the blood to treat SCD. Although there are concerns of off-target effects, it’s a big breakthrough of the year as it opens the possibility of treating SCD using cheap oral pills rather than expensive CRISPR therapy.

To conclude, I expect many discoveries of genetic modifiers of disease causing mutations in the near future. These discoveries will be driven by large-scale sequencing efforts, natural history studies of disease mutation carriers, and importantly, innovation in phenotyping methods.

Drug development is a risky business. Less than 1 out of 10 drugs that enter clinical development succeeds (success defined as getting an FDA approval and entering the market). Because of this extremely low success rate, the cost of developing a drug has skyrocketed. The expectation is that to land on one successful drug, you should work on at least 10, to increase your odds of succeeding. And landing on that one success will offset all the money wasted in developing the failed drugs. This current business model demands huge investment to run a drug business, often in the range of billions of dollars. In a such a scenario, strategies that can push the base success rate beyond 10% will prove highly valuable for drug companies. One such strategy is to embrace human genetics.

Recently I contributed a 'Views and News' article to GEN Biotechnology journal on invitation from the journal's executive editor Kevin Davies. The commentary was on an article published recently in Nature by Matt Nelson and colleagues on the value of human genetics to improve drug success.

Nelson is well known in the field for his landmark article in Nature Genetics from 2015. It was the first empirical demonstration that human genetic evidence can increase the odds of drug success. Through a systematic analysis of public and proprietary databases, Nelson et al. arrived at the conclusion that prioritizing drug targets based on human genetics evidence can nearly double the success rate. The paper was a hit. The impact it had on both industry and academia was beyond Nelson's or anyone's expectation. It became an essential reference in investment pitches made to biotech VCs. It has probably influenced many big investment decisions in the past decade.

Nelson and colleagues' recent work is a follow-up of their 2015 analysis. The core idea of the analysis in both their articles is to curate a list of drug programs that are in various stages of development and evaluate how often programs backed by human genetics succeeded compared to others. It's important to understand what, according to the authors, qualifies as a "human genetics evidence". If the gene corresponding to the drug (that is, the drug works by targeting either the gene or its product) has any evidence of association with the indication in the literature, it is considered to have a human genetics evidence. If there is no published human genetic literature linking the drug target to its indication, it doesn't mean human genetics favors against the drug. It only means there is no published evidence at the moment, but it might surface later as new studies get published. And that is the motivation behind Nelson and colleagues recently following up on their past analysis using latest data. This work was led by Eric Vallabh Minikel from Broad Institute in collaboration with Nelson and two other industry scientists.

Using a proprietary database of drug programs curated since 2000 by Citeline Pharmaprojects, Minikel et al. made a list of ~30,000 target-indication pairs. Using publicly available human genetics databases, they then made a list of ~82,000 gene-trait pairs. Overlapping these two, they found a little more than 2000 target-indication pairs that also had matching gene-trait pairs, which is a mere 7.3% of the total target-indication pairs. This is their main group, that is, the group of drugs with a supporting human genetic evidence. And the rest of 92.7% of target-indication pairs are their reference group, that is, target-indication pairs with no human genetic evidence yet. Here, absence of evidence doesn't mean evidence of absence. Many of the target-indication pairs in reference group can switch to main group as new studies get published, or new databases get released.

So, what did the authors find? The relative success rate of drug programs in the main group (with human genetic evidence) to move from phase 1 all the way to FDA approval was 2.6-fold higher than the drug programs in the reference group (without human genetic evidence). This effect size is 30% more than their 2015 estimate. This is interesting, why? Because it says that the human genetic discoveries are not saturated yet. As more findings get published, the value of human genetics in predicting drug success will further increase. Perhaps that realization is what motivated Nelson to start his company, Genscience. According to the website, the goal of the company is to "integrate world-class genetic insights into drug discovery and development decisions with differentiated expertise, platforms, and analytics." If you'd like to learn more about Nelson and his new company, check episode 118 of The Genetics Podcast, hosted by Patrick Short (YouTube, Apple podcast).

Clearly, Nelson has pioneered a research path that is inspiring many. Recently, 23andMe published a preprint on a similar analysis. Given the recent unfavorable market situation, the company has made a brilliant move to showcase the value of their in-house genetic database. Following the footsteps of Nelson, 23andMe scientists (Wang et al.) combined the genetic association data based on 15 million 23andMe participants (who volunteered to participate in research) with Citeline Pharmaprojects database and evaluated the success rate of drug programs with human genetics support. They arrive at a similar conclusion as their predecessor—drug programs with human genetic support based on self-reported phenotypes are 2 to 3 times more successful than drug programs without human genetic support. Here, the phrase 'self-reported' is notable. Because one criticism that 23andMe has often received in the early years is the questionable value of self-reported phenotypes. And that is one of the reasons 23andMe has been extensively collaborating with academic researchers to show that self-reported phenotypes can produce reliable genetic results as clinical phenotypes when collected at scale. And proving this was an important requirement for GSK to pay 300 million dollars to 23andMe for an exclusive four-year access to 23andMe genetic database in 2018. 23andMe’s recent preprint and its eventual publication will become a key document in all their future deals with drug companies who would want to incorporate 23andMe genetic data into their drug pipelines.

One of the interesting findings in the preprint is the difference in the impact of rare vs common variants on drug success. Minikel et al. argue that minor allele frequency or effect size did not influence the success rate, and that common variants predicted drug success just as efficiently as rare variants, and small effect size variants predicted drug success just as efficiently as large effect size variants. But Wang et al. show otherwise. When stratified by minor allele frequency, genetic evidence based on rare variants seem to double the impact compared to genetic evidence based on common variants, the same true for effect size. Note, these insights are based on imputed rare variants. 23andMe doesn't sequence their participants, they only genotype and impute the variants based on a reference panel. But when you're imputing variants in millions of participants, even rare variants with modest imputation accuracy can yield accurate results, as the sheer size of the data compensates for the reduced imputation accuracy. For example, if you're predicting a rare variant to be present in 20 individuals, of which only 15 are true carriers, you won't have the power to identify genetic associations. But when you are predicting a rare variant to be present in 200 individuals, of which 150 are true carriers, you now start to pick up genetic associations. With further increase in sample size, you gain more power.

Although one might immediately assume that rare variants perform better than common variants because rare variants have larger effect size, and genes with larger effect size are better drug targets than those with smaller effect size. Remember, here we are not using human genetics to discover targets. We are simply using it to increase our confidence on already discovered targets. For such purposes, effect size shouldn’t matter. Effect sizes are reflection of the severity of the genetic variant. A loss of function variant will of course have a larger effect size than a missense variant or an eQTL variant. But as long as the genetic association establishes the link between target and indication, the effect size shouldn’t matter. At least that is the current understanding.

So, then why rare variants perform better in Wang et al’s analysis? It’s because rare variants precisely pinpoint causal genes at the locus of association, but common variants don’t. Both Minikel et al. and Wang et al. are on agreement on this. Minikel et al. show this by using a metric created by Open Targets called locus-to-gene (L2G) score. L2G score derives information from multiple sources such as distance from the index variant, coding variants in linkage disequilibrium, effect on gene expression (eQTLs), overlap with epigenetic marks etc. and predicts the likelihood of a gene to be causal. The authors show that as the L2G score increases (meaning, as the confidence in the causality of the genes increases), the prediction performance increases. Similarly, Wang et al. demonstrate the same, based on their own V2G score estimates. Further, they also show that success rates are higher when causal genes were assigned based on allelic series (meaning, identifying independent variants in the same locus, all linking to the same gene), rare variant burden tests and coding variants in comparison to other methods such as eQTLs, nearest gene etc. We know that former list are better predictors of causal genes than the later.

So, it is becoming clear that we will be seeing more analyses predicting drug program success with human genetics data periodically in the future, either from Genscience or other companies like 23andMe with their own in-house genetic data. But at the end of the day, we shouldn't forget that all these analyses are retrospective in nature where we draw conclusions on past successes. Importantly, it is a biased analysis as we are looking at only reasons for successes but not reasons for failures. Many in the field argue that the real value of human genetics is not to foretell which drugs will succeed but to predict which ones won’t. However, it’s difficult to test this as it’s difficult to confidently tell that a gene A is not associated with a trait X (there could be many reasons why a variant is not associated with a trait). But it’s easy to tell that the gene A is associated with the trait X. So, we can only empirically reason on past successes but not failures.

An ideal analysis should be a randomized trial where you randomly select two groups of drug candidates. In the first group, prioritize drug candidates for clinical development with insight from human genetics, and in the second group, do the same but without any insight from human genetics. Which group will succeed more? Such an experiment is unrealistic. But human genetics is now becoming an important layer of evidence to prioritize drug programs in many companies. Based on the experiences of such companies, we will know the real world impact of human genetics in improving the drug success rate. When I tweeted about my article recently, John Maraganore, the former CEO of Alnylam, said they achieved 60% success rate (that is 6 times the base rate) when using genetics to prioritize their RNAi drugs. Amazing isn’t it? There is no better time than now to be working and investing in human genetics.

Happy weekend! I was reading about Epstein-Barr virus (EBV) this week, inspired by a Nature paper on the association between a cytokine receptor gene, IL27RA, and EBV infection. Researchers from INSERM in Paris have encountered a handful of children with severe EBV infection, who turned out to be knockouts for IL27RA. The gene encodes one of the two subunits of interleukin 27's receptor (IL27RA stands for interleukin 27 receptor subunit alpha). This is the first time researchers are learning about the role of IL27 signaling in EBV infection.

Genetic defects in immune genes increasing the susceptibility to EBV infection is not new. Scientists have known many genes that when lost result in severe EBV infection. But the case of IL27RA is special.

As you might remember learning in high school or in college that EBV is an oncogenic virus, meaning, it causes cancer. Not in everyone it infects. That would be crazy, as more than 95% of the population gets infected with EBV at some point in their life. Most wouldn't even know they did get infected. The virus can silently enter the human body through exchange of body fluids (often through kissing, hence the name 'kissing disease' for infectious mononucleosis) and reach their home that is B cells and live there for the rest of the human’s life. But a subset of humans, typically those with a weak immune system, develops cancer after EBV infection. EBV can cause different types of cancer: lymphomas, gastric cancer, nasopharyngeal cancer etc.

All previously known inborn errors of immunity linked to severe EBV infection have almost always led to cancer. And this is the first time, scientists are coming across a genetic defect that has resulted in severe EBV infection but hasn't resulted in cancer. Perhaps, these patients could be holding the answers to discovering medicines to prevent EBV-infected, immunocompromised individuals from developing cancer.

The discovery story of EBV

In this short YouTube video, Dorothy Crawford and Alan Rickinson, the authors of the book 'Cancer Virus', tell the story of the EBV discovery. The story begins sometime in the late 1950s in Kampala in Uganda, when an African child caught the attention of a British surgeon, Denis Burkitt, working in a local hospital. Fascinated by the facial appearance of the child, Burkitt examined the child. The child had a swollen jaw. The swelling was caused by multiple tumors. Burkitt had never seen such a case before. He made a note of the case in his personal diary and went on his business. Two weeks later, he noticed another kid on the street with a swollen jaw. This cannot be a coincidence, he felt. He dropped everything and went searching for more such cases around the area and found 38 individuals. In some cases, the tumors grew so big they knocked the teeth out and distorted the face severely. He didn't know what caused the tumor, but he felt it must be the most common tumor in that region of Africa. The occurrence appeared to be geographically restricted to areas with high rainfall and hot weather. He published his observations in the British Journal of Surgery in 1958.

On March 22, 1961, Burkitt visited London Medical School and gave a lecture on his recent work titled “The Commonest Children’s Cancer in Tropical Africa — A Hitherto Unrecognised Syndrome”. Among the audience was Anthony Epstein, a young cancer researcher from Middlesex Hospital in London. As a scientist researching the role of virus in cancer, Epstein got excited learning about Burkitt’s work, particularly Burkitt’s intuition that the tumors may have a viral etiology. He proposed to collaborate with Burkitt and requested him to send biopsy samples of the tumors for electron microscopic studies. It seemed that for almost two years, sample after sample, Epstein and Yvonne Barr, his graduate student who helped Epstein on the project, couldn't find anything. By the time they were almost about to give up, the 26th sample flew from Kampala to London on a Friday in December 1963. It was a foggy day, the Heathrow Airport was closed, and the biopsy sample was sent wrongly to Manchester and as result, the delivery was delayed by 12 hours. It was afternoon by the time the sample reached the lab. When they opened the package, they found that the fluid containing the biopsy was cloudy. They almost threw it away as the sample appeared to be contaminated with bacterial infection. But when looked through a microscope, they realized that the cloudiness was because of free tumor cells floating around in the medium. They fished out the tumor cells and dropped them in a Petri dish containing cell culture medium. The tumor cells began to grow. In about six weeks, they had a fully grown tumor cell culture containing malignant lymphoid cells. Sometime in February 1964, Epstein examined the grown tumor cells using electron microscope. He couldn't believe his sight. He switched off the microscope and walked around the block for 20 minutes. Then returned to the lab and looked at it again. It hit him what he saw was real: small virus particles lurking inside the tumor cells. And that's how the first human tumor virus was discovered, says Crawford and Rickinson.

EBV and multiple sclerosis

Sixty years have passed since the EBV discovery, still the virus surprises scientists year after year. Just a couple of years back, there were two studies, one published in Science and the other in Nature, which revisited the long-suspected link between EBV and MS and gave a compelling proof on the causal role of EBV in MS.1 In the Science paper, using 20 years longitudinal health data of US military recruits, researchers from Harvard T.H. Chan School of Public Health reported an epidemiological analysis, proving the causality of EBV's association with MS with near certainty. In the Nature paper, Stanford researchers reported a follow-up analysis on the EBV's association with MS, detailing the molecular mechanisms through which EBV infections might lead to MS. The authors found that one of the EBV antigen is a doppelgänger of a central nervous system (CNS) protein called glial cell adhesion molecule (GlialCAM). As a result, CNS glial cells become victims in the war between the body's immune system and EBV, leading to autoimmune-mediated myelin destruction. An mRNA vaccine against the viral antigen EBNA1 that mimics the GlialCAM is currently trialed as a means to prevent MS after EBV infection.

EBV and IL27 signaling

A bulk of the research on EBV over the past half a century went into understanding immune signaling pathways that are involved in the EBV infection and the development of complications such as cancer. Naturally occurring genetic mutations in humans, resulting in a breach in the immune defense against EBV has helped scientists to trace the EBV-related immune signaling pathways. Below is a illustration of the cross talks between EBV-infected B cells and T cells at initial and later stages of immune defense. The proteins highlighted in red are the ones identified to play key roles in the immune response against EBV infection through inborn errors of immunity. The IL27RA adds to this list of genes.

The authors of the current Nature paper, researchers from the INSERM in Paris, studied three children with severe acute EBV infection, unusual for their age. EBV infection in children are usually asymptomatic. Clinical and lab investigations suggested that the children might have some immune-related genetic defects. As suspected, the authors found recessive deleterious mutations in an immune-gene, IL27RA, encoding the receptor of cytokine interleukin 27 (IL27). Two of the children were homozygote and one, compound heterozygote. Referencing the variants with public databases, the authors found that the variants were not reported before, except one, rs201107107, which occurred in relatively higher frequency (MAF=0.7%) in the Finnish population. There were 15 homozygotes of this variant in the Finngen. This motivated the authors to collaborate with the Finngen team to find out if any of these 15 homozygotes have got severe EBV infection. Two of them turned out to have been hospitalized for infectious mononucleosis, which is 50-fold enrichment compared to the non-carriers.

Through in vitro experiments in the blood cells from the three patients, the authors found that T cells of the patients indeed do not express IL27 receptors or up regulate IL27RA expression upon activation, unlike healthy controls. Stimulating the T cells with IL27 did not activate the downstream cell signaling events either. In the paper, the authors present further a wide range of functional experiments, highlighting the role of IL27 signaling in the T cell response to EBV infection. The bottom line: the three patients with IL27RA mutations have helped scientists learn about the role of IL27 signaling in the EBV infection. Particularly, its possible role in EBV-mediated oncogenic transformation of B cells. It seems the same receptor that is required for T cells to respond to EBV-infected B cells to fight against infection, is also required for T cells to transform the EBV-infected B cells into tumor cells. As a result, even though loss of IL27RA results in severe EBV-infection, it doesn't further lead to cancer development. This opens up the possibility of targeting IL27 signaling pathway to treat or prevent cancer in individuals at high risk of developing severe EBV infection, such as those with weak immune system. But it also makes me wonder if IL27 signaling plays a role in MS development. If so, it will also open up therapeutic opportunities for the treatment of MS. Time will tell.

I've known about LRRK2 and that many companies are developing drugs to inhibit this gene in the brain to treat PD. But I never realized LRRK2 is a field of its own and hundreds of researchers across the world are religiously studying this one gene since 2004, trying to understand how LRRK2 mutations result in the death of dopaminergic neurons in the brain. It all started with two publications in 2004 in the journal Neuron.

One of the many loci mapped to PD through family-based linkage analysis was PARK8, a large locus spanning the centromere ends of short and long arms of chromosome 12, reported first in 2002. The linkage region spans many millions of base pairs and hold more than 100 genes. The locus was originally mapped in a Japanese family, but soon in families in the Europe, UK and US. The PD-causing gene was dragged out of the gene-dense forest in 2004 by two research groups, one led by Thomas Gasser from the University of Tübingen in Germany (Zimprich et al. Neuron 2004) and the other, Andrew Singleton from the NIH in the US (Paisán-Ruíz et al. Neuron 2004)

Thanks to a series of families from Basque region of Spain, the disease markers that segregated in these families helped Singleton and colleagues to narrow down the disease locus from ~12 Mbp containing 116 genes to ~2 Mbp containing 11 genes and eventually to the causative gene, LRRK2. The gene encodes a gigantic protein made of 2,482 amino acids. Honoring the Basque families that helped clone this gene, the authors named this gene as dardarin, based on the Basque word 'dardara', meaning tremor.

It was the first time the cell signaling field was hearing about this gene. The researchers didn't know what kind of protein LRRK2 was making. Based on the amino acid sequence, they noticed that there were many domains, one of which was a kinase domain where many of the disease mutations appeared to cluster. The gene encoded a kinase protein (that is, it is capable of enzymatically adding a phosphate group to another protein or to itself), and the kinase activity is perhaps at the center of the biology of LRRK2-mediated PD, the researchers guessed. They proved the same in two years. Removing the kinase domain of LRRK2 eliminated its neurotoxicity. So, inhibiting the kinase activity of LRRK2 is the key to treating PD, at least the one caused by LRRK2 mutations. This realization marked the beginning of the quest that many researchers, particularly the ones in the cell signaling research, would dedicate their careers to understanding the molecular mechanisms underlying the link between LRRK2 and PD.

On one side, drug companies focussed on developing safe, CNS-penetrant LRRK2 inhibitors to treat PD. Denali therapeutics is at the forefront of this area; its drug candidate is currently in late stage clinical trials. On the other side, cell signaling researchers focussed on tracing the signaling pathways that link LRRK2 to neurodegeneration. I learned much of this from listening to the lectures (2017, 2020, 2021) of Dario Alessi, renowned PD researcher from the University of Dundee in the UK. Before 2004, Alessi was working on the protein kinase biology in the disease areas like hypertension and cancer. But then the news about LRRK2 hit Alessi like a calling and as a result, he spent the next 20 years studying LRRK2.

In the early years of LRRK2 research, Alessi and others in the field were trying to answer one question. Every protein kinase must have a substrate. What is the physiological substrate of LRKK2? This seemingly simple question costed more than a decade for the researchers to arrive at the answer. Through mass spectrometry experiments, studying the phosphorylation of profile of hundreds of proteins in the presence and absence of LRKK2 activation in cell cultures and in the brain tissue of animal models, researchers found in 2016 that the substrate of LRRK2 is a family of proteins called Rab (Ras-related protein in brain).

Rab proteins are GTPases, and there are 70 Rab GTPases known to exist in the humans, performing a wide array of molecular functions. A subset of these interact with LRRK2 and contribute to neurodegeneration. Researchers have short-listed a few that are phosphorylated by LRRK2 such as Rab8, Rab10 etc. and studying those in detail will help understand how their phosphorylation leads to neuronal death. Phosphorylation of these Rab GTPases seem to impact a wide range of cellular processes such as protein trafficking, ciliogenesis, etc. Understanding which of these are in the causal pathways to neurodegeneration is critical to inform therapeutic development. Notably, none of these downstream candidate Rab GTPases have been linked to PD via human genetics.

After figuring out the physiological substrates of LRRK2, researchers wondered what's upstream of LRRK2. It turns out LRRK2 is mainly a cytosolic protein, but Rabs are membrane-bound proteins. The Rabs are bound to the plasma membrane but also to the membranes of cell organelles like ribosomes, lysosomes, endoplasmic reticulum and Golgi. So, something has to reel the LRRK2 proteins from cytosol to the membranes where Rabs hang out. When the LRRK2 are brought to the membranes in proximity to the Rab proteins, it gets activated. So came the next big question: what recruits LRRK2 to its site of action? Researchers feel that the answer to this question will help understand the molecular mechanisms of LRRK2 activation, which in turn will help design safe and effective drugs to inhibit LRRK2 activation to treat PD. So far, two proteins, Rab29 and VPS35, were found to contribute to LRRK2 activation from upstream.

The RAB29 gene (also known as RAB7L1) is located in the PARK16 locus (one of the loci discovered by early GWAS studies) and is believed to be the causal gene, though no strong rare variant associations seem to exist in the literature, except for a few candidate gene studies. However, Rab29 protein has been extensively studied, particularly by Alessi's research team, in the context of LRRK2 activation. Rab29 binds to LRRK2 through one of LRRK2's non-catalytic domains called armadillo domain. Below is a beautiful electron microscope-based structures of LRRK2 bound to Rab29 in different configurations (monomeric, dimeric and tetrameric) reported in Science last year.

The other protein that is shown to activate LRRK2 from upstream is VPS35, which encodes a component of a multi-protein complex called retromer complex that transports used and worn out proteins by transporting them from endosomes to the Golgi for recycling. Mutations in VPS35 causes autosomal dominant Parkinson's disease, and it turned out LRRK2 is over-activated when VPS35 is mutated. The VPS35 protein seem to negatively impact LRKK2 activation. Monocytes and neutrophils of PD patients with VPS35 mutations show excessive phosphorylation of Rab10 (LRRK2’s substrate), suggesting LRRK2 may be mediating the neurodegeneration in PD patients with VPS35 mutations. The exact mechanism through which VPS35 contribute to LRRK2 activation is yet to be understood.

With the recent ExWAS of familial PD, Rab32 now joins the Rab party. Studying the exome sequences of 2,184 familial PD cases and 69,775 controls, an international research team discovered a missense variant, Ser71Arg, in RAB32 that increases the risk of PD by almost the same effect size as some of the severe LRRK2 risk variants. And it is seen in 0.7% of the familial PD cases. The gene also came up as the top risk gene in an independent study that focussed specifically on the Rab mutations in a small sample of familial PD cases. Probably, we will read more reports replicating RAB32's link with PD in the near future as many will now scan through the exomes of their PD cohorts, looking for Ser71Arg mutation.

Rab32 is already known to directly interact with LRRK2 through its armadillo domain, just like Rab29. The two studies that report the RAB32 mutation also show some evidence that LRRK2 and its downstream Rab10 are both hyperphosphorylated in the blood cells of patients with RAB32 mutation. Interestingly, the missense variant position 71 is the only phosphorylation site in RAB32. It's not clear if Rab32 is upstream or downstream of LRRK2 in the PD pathology (looks like upstream). But we will know soon, as many in the LRRK2 research have probably already started investigating RAB32.

The Ser71Arg mutation appears to be a gain of function mutation. Many questions to answer. Will Rab32 inhibition help treat PD in these patients? If Rab32 plays a key role in LRRK2 activation, then, will PD patients with LRRK2 mutations benefit from Rab32 inhibition? Conversely, will PD patients with RAB32 mutation will benefit from LRRK2 inhibition. This would expand the patient population for LRRK2 inhibitors. Clearly, the discovery of Rab32’s link with PD has opened many exciting research questions to explore.

Happy Weekend! Recently, I've been reading about the origin story of lab mice in an essay, The Mouse as a Microscope, by Alex Telford1.

Asimov Press

Origins of the Lab Mouse

Today we launch Issue 03 of Asimov Press. Read our full Editors’ Note and preview upcoming articles by clicking here. Read this article on our website by clicking here…

Read more

2 years ago · 27 likes · 4 comments · Asimov Press and Alex Telford

As a human geneticist working in the drug development field, I, like many others, have often complained about mice being a poor animal model to study the complex human physiology. But I've also appreciated the big roles those small mammals have played in biomedical research, particularly in human genetics. A few days ago, I tweeted about a new discovery of a rare monogenic cause of childhood onset systemic lupus erythematosus (SLE)—gain of function mutations in UNC93B1, encoding a toll-like receptor (TLR)-trafficking protein—published in Nature Immunology. The discovery reminded me of another post from two years ago on the discovery of a gain of function mutation in TLR7 as a monogenic cause of childhood onset SLE. With Alex's story still fresh in my mind, an aspect of both the studies kept coming back to me: the roles lab mice have played in establishing the links of TLR7 and UNC93B1 with SLE.

SLE is an autoimmune disease. Individuals with SLE produce autoantibodies against their own DNA, commonly known as antinuclear antibodies (ANAs). Under normal physiological conditions, DNA never leaves its home, the nucleus. When it does, it causes trouble. It alarms the innate immune system. Humans have evolved by fighting against pathogens like viruses for thousands of years. Our immune system is pretty good at spotting a pathogen intruder. When a virus enters a human cell and start replicating, their nucleic acids are readily sensed by a special family of badass proteins, TLR7, TLR8 and TLR9, that live in the endosomes. Activation of TLRs leads to consequences such as cytokine storm, B cell activation etc. to fight against the intruder. In patients with SLE, there is constant activation of TLRs without any viral infection, either because of frequent release of nuclear DNA into the cytoplasm or abnormally sensitive TLRs.

Presence of ANAs is diagnostic of SLE, which is known for a long time. ANA test is highly sensitive but not specific for SLE. As Derek Lowe puts it in one of his In the Pipeline articles, "if you are negative on an ANA, you basically don't have lupus. But if you're positive, you could still have something else, because that's a common positive in a number of autoimmune diseases". The groundbreaking discoveries of the TLRs' role in pathogen nucleic acids sensing in the early 2000s gave the first clues on the link between TLRs and SLE. It biologically made sense that TLR over activation could cause SLE. But the real-world evidence wasn't there to prove this causal link, though there were some hints. You see, TLR7 and TLR8 are on the X chromosome, and SLE is more almost 10-times more common in females than males. Scientists always thought that these two facts are not unrelated. Is it possible that the increased prevalence of SLE in females could be due to increased expression of TLR7 in the immune cells? We know the answer to that question today, that certain immune cells in females escape the random X inactivation and double the dosage of their X chromosomal genes, including TLR7. But in the early 2000s, scientists weren't sure. They were just beginning to connect the dots.

What if we could stitch an extra copy of TLR7 into the Y chromosome, will that increase the risk of SLE in males? Well, it turned out that is what exactly happened when scientists crossed B6 female mice (B6 or C57BL/6 is the most commonly used inbred mouse strain) with SB/Le male mice (SB/Le is a special strain used to study autoimmune diseases). This resulted in a hybrid that showed extreme autoimmunity. The animals spontaneously developed a lupus-like disease. The reasearchers at the National Institute of Allergy and Infectious Diseases in NIH named their creation Yaa (Y-linked autoimmune accelerator). It appeared that the crossing of B6 and SB/Le strains resulted in some genome exchange that pushed the autoimmune risk of SB/Le male mice through the roof. There was a clue that X chromosome from the B6 strain has something to do with it. Because Yaa mice came only from an SB/Le dad and a B6 mom. But not the other way around.

When researchers took a closer look at the Yaa’s genome, they stumbled upon something wild: an uninvited guest, a tiny piece of X chromosome containing the TLR7 gene, lurking inside the Y chromosome. So, the Yaa mice have got two copies of TLR7: one of their own in the X and another in the Y as a result of a translocation. The TLR7 duplication amplified the already autoimmune susceptible genome derived from the SB/Le strain. This spontaneous mutation provided a causal evidence that TLR7 over activation results in SLE. Conversely, when TLR7 is deleted in lupus mice model, scientists found, the autoimmunity is dampened.

Despite the overwhelming lines of evidence, no strong human genetic findings have surfaced connecting TLR7 gain of function with SLE susceptibility for a long time. The first such report was in 2022, which I have summarized in a Twitter thread. Scanning through the genome of a 7 yr old girl with a refractory, early onset SLE, a research team in Australia (Brown et al. Nature 2022) caught a de novo missense variant in TLR7 within a highly conserved sequence that encodes the ligand-binding part of the TLR7 protein (TLR7’s ligands are single stranded RNAs). Introducing this variant in mice triggered autoimmunity. If you read the paper, you'll appreciate how this single variant has opened a window for the immunologists to peek into the world of a SLE pathogenesis, revealing tons of insights. But what fascinated me most is that the authors knew where to look at. Thanks to prior knowledge on the role of TLRs in SLE that came mainly from mice studies.

The memory of the TLR7's story is what made me excited when I came across the Nature Immunology paper by a research team in China (Al-Azab, Idiiatullina, et al.) on gain of function mutations in UNC93B1 causing childhood-onset SLE. The beauty of these discoveries is that they are all interconnected, and discovery of one gene exerts a domino effect on the other. If too much TLR7 causes SLE, other proteins upstream or downstream of TLR7 signaling should have genetic links to SLE. One such protein is ‘unc-93 homolog B1’ encoded by UNC93B1.

Until the early 2000s, UNC93B1 was one of those genes encoding a protein with a weird name and an unknown function. Then born a mouse with the ‘3d’ mutation. Both of its parents were created using gametes bathed in a chemical called N-ethyl-N-nitrosourea, a powerful mutagen2. One of the hundreds of random chemically induced mutations hit UNC93B1, resulting in an immunodeficiency phenotype. It turned out that one stone killed three birds—TLR3, TLR7 and TLR9, opening the doors to all sorts of pathogens. The immunologists at The Scripps Research Institute who discovered this UNC93B1 mutation decided to call it ‘3d’ as it caused 3 defects in the immune system. Studying the ‘3d’ homozygous mice revealed the critical role unc-93 homolog B1 played inside the immune cells. It trafficked the TLRs from the endoplasmic reticulum to their site of action, the endosomes. Complete loss of UNC93B1 resulted in a selective immunodeficiency against intracellular pathogens. In the same year, the human counterparts of ‘3d’ mice were encountered by a research team in Paris (Casrouge, Zhang, Eidenschenk, Jouanguy, et al. Science 2006): two French children born from of a consanguineous marriage who were knockouts for UNC93B1 contracted a herpes simplex virus 1 infection in their brains.

Since its discovery, UNC93B1 has been extensively studied in the immunology field in the context of TLR signalling. In vitro mutagenesis screen has shown that gain of function mutations in UNC93B1 can amplify TLR signaling. However, no naturally occurring gain of function UNC93B1 mutations in humans were known. But it was only a matter of time. To find such mutations, all one has to do is look for them in the right individuals, which is exactly what the research team at the Guangzhou Medical University in China did.

Scanning through the genomes of a childhood-onset SLE cohort, the team landed on the UC93B1 mutations, which they suspected to cause SLE. The mutations were in the part of the protein that binds specifically to TLR7. Introducing the mutation in mice triggered autoimmunity, confirming their intuition that these are gain of function in nature. The interesting part of the findings is that these mutations appear to be exclusive to East Asian ancestry. Particularly one of the mutations showed a gradient along the longitude with the highest frequency occuring in Southern coastal Han Chinese population. This is an important observation as this would mean that drug targeting TLR7 might be of value specifically to the patient communities from this part of the world, an example of personalized medicine in the context of ancestries. The good news is TLR7 antagonist do already exist that can be used to treat SLE in these patients.

To conclude, in both of the above stories, I wanted to emphasize the role lab mice played in kickstarting the research fields that led to an immense understanding of TLR signaling and how they are disrupted in SLE. If you crave for more inspiration about lab mice, I recommend Alex’s essay The Mouse as a Microscope.

Happy Friday! One of the talks at the ASHG 2023 blew the minds of the audience. It was a talk on single cell sequencing of brain tissue from Huntington's patients, presented by Bob Handsaker from Steve McCarroll's group at the Broad Institute in Boston. I didn't grab the whole story during the presentation. But the parts that I heard and the final revised disease model that Bob presented felt groundbreaking, and I couldn't contain my excitement. I tweeted about it that night, and many agreed with what I felt. I spoke about it with Patrick Short (who was as excited as I was about this work) in the 2023 year-end episode of The Genetics Podcast. However, I feared that I may have overblown the impact of the work prematurely, and the balloon would burst someday when the actual preprint comes out. Like many others, I've been waiting for the preprint to drop.

Steve's team finally posted their work in the medRxiv a week ago. Having now read the paper fully, I can say with full confidence that the hype that Bob's ASHG presentation received is well deserved. The work described in the preprint is as just as impressive as I imagined. The work has already garnered a lot of excitement in the community. Here is a tweet from Mark Daly, a reputed, world famous scientist in human genetics. I don’t remember hearing such a praise from Mark on any work before.

What’s known?

It's important we first understand what's already known to appreciate what's new. Huntington's disease is caused by expansion of a microsatellite mutation--CAG repeat--located in the exon 1 of HTT gene in chromosome 4. The number of CAG repeats in the general population range between 15-30, whereas in Huntington's patients it is between 36 and 55.

The CAG repeats in the HTT gene transcribe and translate into poly-glutamine chain in the Huntingtin protein; the HTT is expressed in the fetal and adult brain, among other tissues. The huntingtin protein with an elongated poly-glutamine tract is neurotoxic. It erodes the neurons, particularly the neurons of the striatum, over time in the Huntington's patients, resulting in motor, cognitive and psychiatric symptoms.

Huntington's is a dominant disease. Elongation of only one allele is sufficient to fry the brain. A second hit doesn't have any addition impact on the disease course. There is a long latency period. The age of onset is around late adulthood, typically around 40 years. It takes decades before the symptoms manifest. Though we don't know what causes this latency, we know that the age of onset correlates with the CAG repeat length; longer the repeats, earlier the disease onset.

The CAG repeats in HTT expand across generations (in the germline) but also within a generation (in somatic cells). The phenomenon of expansion across generations is described as "anticipation", where the disease becomes severe in subsequent generations. The germline expansion occurs as a consequence of DNA polymerase slippage, resulting in strand loops. The DNA mismatch repair proteins while attempting to fix the loop end up adding more DNA base pairs than what was originally present, resulting in repeat expansion.

The somatic expansion mechanism is less understood than the germline expansion, particularly in the post-mitotic cells such as neurons. One theory is the expansion is a consequence of limitations of the DNA repair system to properly fix the DNA strand loops that happen during gene transcriptions, particularly genes with repetitive regions like HTT. Instead of excising the lengthy strand and synthesizing a matching short strand, the confused DNA mismatch repair proteins do the opposite: excise the short strand and synthesize a matching long strand. This theory is strengthened by the fact that genetic variations in mismatch repair genes modify the penetrance of HTT repeats.

What’s not known?

• Of all the brain regions, why striatal neurons are specifically affected in Huntington's disease?

• Clinically, a threshold of 36 repeats in the germline is considered pathogenic. But at the molecular level, what is the threshold for pathogenicity?

• It's known that somatic repeat expansions are contributing to neurodegeneration. But how? Are somatic expansions the primary mediators of neurodegeneration, or they simply modify the severity and rate of neurodegeneration?

• Why is there a long latency? Is it because the toxic huntingtin proteins slowly kill the neurons, or they wait for a period of time (perhaps to become long enough to hurt) before starting to devour the neurons?

• At what time point, the neurodegeneration becomes completely irreversible? Is there a therapeutic window of opportunity to reverse or halt the disease progression?

The current paper

The current work by Handsaker et al. doesn't answer all the unknowns. But they add major insights into the missing pieces of the Huntington's puzzle through single cell sequencing of brain tissue from ~60 Huntington's patients and ~50 controls. I was never a big fan of single cell sequencing technology, as I always felt that it was hyped more than it delivered. This is probably the first time, I was able to truly appreciate the value of this technological advancement.

In a nutshell, by studying the mRNA transcripts of individual brain cells of patients with varying degrees of neurodegeneration, the authors were able to model the molecular disease course of Huntington's disease.

Loss of striatal neurons

Firstly, the authors demonstrate the well known selective death of striatal projection neurons (SPN) in the brains of Huntington's patients. Look at the dramatic reduction of SPN proportion in the caudate nucleus (most affected part of the striatum) of the Huntington's patients in below plot.

Inherited CAG repeats vs SPN loss

Next, the authors demonstrate the relationship of SPN loss with age and CAG repeat length. They use a standard metric called CAG-age-product (CAG) score that reflects the cumulative exposure of a Huntington's patient to expanded CAG repeat. It is calculated by multiplying age with germline CAG repeat length minus 33.66. The relationship between CAP score and SPN loss in the below plot captures both the latent period (CAP< 300) when the SPN loss is only moderate and disease onset (CAP> 400) when there is rapid neurodegeneration. At advanced stages (CAP>600), most of the SPNs (>80%) are lost.

Somatic repeat expansions

Having established the known relationship between SPN loss and germline CAG repeats, the authors turn their attention to somatic repeats. The authors quantify the CAG repeat length directly from the HTT transcript in each of the brain cells. The authors found extensive somatic expansion of HTT CAG repeats in the SPNs, but only mild expansions in the other cell types. This is an important finding, and it suggests that the selective death of striatal neurons is not because they are more vulnerable to polyglutamine toxicity than other cell types, but because huntingtin protein with extremely long polyglutamine tract is produced only in the striatal neurons. Should such toxic huntingtin proteins are produced in other cell types, they also will likely die, which is what we see in Huntington's mice models with extra-long CAG repeats in the germline.

Armadillo distribution

The authors found that not all the SPNs showed similar expansion. While the bulk of the SPNs showed a moderate expansion (around 20–30 repeats more than the germline), a tiny proportion of the SPNs showed extreme expansions (around 100-500 more than the germline). The two groups form a characteristic distribution shape, which the authors compare to an armadillo. They call the body of the armadillo, phase A and the tail, phase B. It seems earlier studies failed to capture the extremely elongated CAG repeats because of technical limitations of the PCR methods.

Repeat expansion vs gene expression

This is the most exciting (and innovative) part of the work. The authors compare gene expression changes between different repeat lengths. Thanks to somatic mutations, each person has their own allelic series of diverse CAG repeat lengths. The authors leverage this fact to perform a within-person comparison of gene expression changes (as a read out of pathogenicity) across different CAG repeat lengths and avoid all the confounders that will arise when comparing between persons. That's a brilliant idea! The results of this analysis are the core findings of the paper.

The authors find repeat expansions of up to 150 repeats had no major impact on the gene expressions. But beyond 150 repeats, there is a dramatic impact, distorting the expressions of hundreds of genes.

Wait, here is the most interesting part. They found that the gene expression changes are highly reproducible across individuals. Look at this correlation plots. Incredible!

Next, the authors trace how the gene expression distortion worsens with increasing CAG repeats. Based on their observations, they define different phases. Remember, we have already seen phase A (the body of armadillo with <100 repeats) and phase B (the tail of armadillo with >100 repeats). The authors now add more phases to the tail.

The phase B ends at 150 repeats, the magic number below which there is no much harm, but once past 150, things go south quickly. The authors bucket the neurons with repeats >150 in three phases: C (Continuous escalation), D (De-repression) and E (Elimination).

During phase C, with expanding repeats, there are continuous changes in gene expressions; they get more and more chaotic to a point where neurons lose their cellular identity. The gene expression patterns of individual cell types are like fingerprints. If you're familiar with scRNA-seq paper, you'd remember seeing tSNE plots showing clusters corresponding to different cell types. Such identities are lost in phase C. At this point, you can no longer tell apart striatal neurons from other neuronal types, or even from a non-neuronal cell type.

Next comes the phase D, the de-repression phase. Now, zombies begin to crawl out of their graves. Genes that are normally disallowed in the striatal neurons start expressing because of the loss of repression. At this point, the CAG repeats have surpassed 350 and the clock is ticking much faster. The authors explored what sort of genes are de-repressed. They found they are mainly transcription factors and noncoding RNAs normally expressed during early embryonic development but not in the adult neurons. It’s as if the neurons are aging in reverse all the way to how they were inside the womb before disappearing forever.

Then comes the final phase, phase E, the elimination phase1. The neurons now have reached the end of their life and so does the patient. Almost all the SPNs are lost and the caudate nucleus is fully atrophied, ultimately resulting in the patient's death.

ELongATE neuropathology model

Putting all five phases together, the authors propose a new model of neuropathology of Huntington’s disease called “ELongATE” (extra-long repeats acquire toxic effect) where the striatal neurons go through five phases before their death.

The slowly, capriciously ticking DNA clock.

During the phase A, the HTT CAG repeats expand slowly and aynchronously. The authors write “We estimate that an SPN takes 50 years (on average) to expand from 40 to 60 CAGs, then another 12 years to expand from 60 to 80, …”.

“Asynchronous expansion” is the key term. The repeat expansion is a stochastic process. So, rate of expansion of one neuron is completely random compared to an another. Yet they all expand slowly and spend >98% of their lifetime in the phase A (armadillo body). The authors compare phase A to “a slowly and capriciously ticking DNA clock.”

The rapidly, predictably ticking DNA clock.

When the neurons reach phase B (80 to 150 repeats), their repeat progression become predictable. As the CAG repeats become longer, their chances of undergoing new expansion mutations increase. As the neurons step inside phase B their life course becomes predictable, and their pending years of life can be timed. The authors compare phase B to “a rapidly, predictably ticking DNA clock”.

Interestingly, even during the rapidly progressing phase B the neurons still seem to function and so, there are no symptoms. Only after entering phase C, the neurons start eroding (as indicated by the gene expression distortion), resulting in symptoms. And quickly they pass through phases D and E.

Therapeutic window

Based on their analysis, the authors predict that there is a long therapeutic window of opportunity during which any interventions that can halt the repeat expansion can be effectively made to delay the disease. The authors discuss that even after the symptoms onset, the disease progression could be slowed down by rescuing the bulk of the phase A neurons from entering the next phase.

Stopping repeat expansion is the key

Based on the findings, it looks like the therapeutics for Huntington’s should be focussed on stopping the repeat expansion. Past drug development efforts focussed on reducing the toxic huntingtin protein, which didn’t work. Both human genetics and animal models have shown that interfering with mismatch repair system slows down the disease. Likely, many companies are currently working on targeting mismatch repair genes to treat Huntington’s disease. If such therapeutic designs turn out safe and effective, we will be seeing a newer generation of miracle drugs for not just Huntington’s but many other repeat expansion-related neurodegenerative disorders.