The Edges of the Map

What I am talking about here is not within my field. I work near it, computational genomics, close enough that the vocabulary is familiar, but the genetics of brain disease specifically is something I've spent the last few months reading about, not doing. During my junior year, I took Advanced Algorithms in Computational Biology and Medical Bioinformatics with Professor Sorin Istrail. One of our projects was to present a GWAS paper of our choosing, and that assignment ended up pulling me much deeper into this literature. But none of what follows is a verdict. It's me just thinking out loud about a handful of papers that have been sitting uncomfortably with me, and that I haven't been able to put back down.

Here's what pulled me in.

In 2007, four separate papers landed in the same journal, all pointing at the same genetic variant linked to coronary artery disease. It read like a proof of concept. Scan enough genomes, compare enough people, and the hidden architecture of disease would surface on its own. That was the promise of genome-wide association studies (GWAS), and it was a genuinely amazing idea. Build a map of inherited risk, one variant at a time, and let the biology draw itself.

The map got built. Hundreds of diseases, enormous scale. The problem isn't that it doesn't exist. The problem is that in a lot of places, we can't read it.

For plenty of common diseases, that's fine, or at least steadily improving. GWAS has flagged thousands of risk loci and pointed toward real pathways. But for the brain, the neurological and psychiatric disorders, the distance between what was promised and what arrived is wide. And the more I read, the more I started to think the reasons for that gap aren't the reasons you would think.

The basic disappointment is this. GWAS can tell you that a variant near some gene is associated with, say, schizophrenia. It cannot tell you what's actually broken. There's a paper by Connally et al. that makes this concrete: even when the culprit gene is known because it causes a rare, clean Mendelian version of the same disease, the standard methods can only link the GWAS signal to that gene's expression about eight percent of the time. Eight percent. When we've been handed the answer key.

The variants mostly land in regulatory DNA, the switches, not the genes themselves. And we usually can't tell what they're switching, probably because the effect only appears in a particular cell state, or a particular developmental window, or under some perturbation nobody's thought to apply. There's a name for this: missing regulation. It's a different problem from the one everyone actually talks about.

The one everyone talks about is missing heritability, the fact that the variants GWAS finds explain only a slice of why these disorders run in families. And it turns out part of that famous gap was a measurement artifact. Golan et al. showed that the standard estimation method quietly underestimates heritability in case-control studies, for a statistical reason baked into the math. Correct for it, and common variants explain closer to 60% of heritability, on average.

But "on average" is doing a lot of work in that sentence, and the brain is exactly where the average falls apart. A polygenic risk score for schizophrenia explains something like 7 to 10% of the variance in European populations. Less, elsewhere. Part of that failure has an obvious explanation, one the field has been slow to fully reckon with.

We didn't study everyone. We studied one kind of person. Jonson et al. looked across neurodegenerative disease GWAS and found that 82% of them were majority-European. For frontotemporal dementia and Lewy body dementia, the number of non-European or multi-ancestry studies was zero. Not few. Zero.

The biology itself doesn't hold still across populations. APOE4, the single biggest known risk variant for Alzheimer's, has a weaker effect in people of African ancestry and a stronger one in Japanese populations. C9ORF72, a major ALS gene, turns up in about 7% of Europeans and 0.3% of Chinese populations. We assembled our whole picture of these diseases out of one slice of human genetic variation, and then quietly assumed the picture was the whole thing.

So it shouldn't be a surprise that the tools we built out of that picture fail the people who weren't in it. Polygenic risk scores are where GWAS reaches the clinic, and they lose somewhere between 20 and 80% of their predictive accuracy when you carry them across to non-European populations. For Alzheimer's specifically, the accuracy gap between Europeans and people of African ancestry runs to roughly four and a half fold.

This isn't only a story about who volunteered for which study. It's structural, and it's almost physical. The SNP arrays, the actual chips that read the genotypes, were designed from mostly European sequencing data. The bias is soldered in at the hardware level. So you can recruit more diverse cohorts, and still be reading them with an instrument tuned to someone else. Deep learning methods can patch over some of this, and people are building them (something I would love to work on). But a patch over a structural problem is still a patch.

Of all the places for this to land, the brain is the worst one. Brain disorders are already the hardest thing GWAS tries to do: extraordinarily polygenic, heavily shaped by environment, hard to phenotype cleanly, often late-onset. The places where the method is weakest are exactly the places where the representation gap does the most damage. The two failures don't just coexist, they stack on top of each other.

And the stakes aren't symmetric. A biased risk score for cardiovascular disease is a problem worth fixing. A biased risk score for Alzheimer's or schizophrenia, handed to a clinician who's applying it to someone of African or South Asian ancestry, isn't just less accurate. It can point in the wrong direction, with the numbers to back it up.

I'm wary of concluding this with a definitive statement, because I'm genuinely not the person to write one. But the honest math seems to be something like this. More diverse cohorts matter, and aren't enough on their own, not while the arrays stay biased. Whole-genome sequencing would sidestep the array problem entirely, except the cost is out of reach at scale for most of the countries that would benefit most. Multi-ancestry methods, things like PRS-CSx and GAUDI, are showing real gains. And the field is moving: more than 65% of the non-European loci for these diseases were found after 2020. But moving toward something and getting there are different, and I don't think we're there yet.

The organ we understand least. The populations we've studied least. I don't think those three facts are independent, and I don't think that's a coincidence.

So I've come around to thinking the question was never really whether GWAS was the wrong tool. It's a good tool. The question is whether we used it carefully enough, on a wide enough sample of actual humanity, to have earned the confidence we've placed in what it told us.

Yesterday, the NIH announced that its All of Us Research Program has become the largest integrated genomics and health database in the world, with over 86% of its 747,000 participants coming from communities historically underrepresented in biomedical research.

It is, genuinely, a reason for optimism.

The gaps in the map aren't blank.

They're full of people.