Why human genetics and AI are a great match
CEOs of AI labs are claiming that AI is going to solve climate change and cure cancer. The technology they are developing is already changing our lives with chatbots and deep fakes, but climate change and cancer are still waiting for AI to solve them.
With all the hype around AI, it’s easy to be cynical about it. But we have seen AI doing truly remarkable things. A big part of the excitement about AI has to do with recent progress in biomedicine, especially around proteins. It’s been less than 4 years since AlphaFold2 solved a huge part of the protein folding problem (which just won a Nobel prize), and there’s been a lot of progress since.
One of the technologies I’m most excited about (and personally involved in) is protein language models. They work in a similar way to text language models like ChatGPT: given a sequence, the model tries to predict what comes next. But instead of predicting the next word in a text, protein language models predict the next amino acid in a protein sequence. Like with text, this produces very capable models.
It turns out that protein language models learn a lot about the structures that proteins fold into and what proteins do (their function) without needing to see this information explicitly. They learn a lot about proteins simply because it’s useful for making informed guesses about what comes next in a protein sequence. For example, knowing the structure of a protein can tell you that certain amino acids don’t fit in specific positions.
Language modeling in genomics is in some ways even more promising than language modeling in text. Text language models are trained on text written by people, with the purpose of teaching the models what we already know. But genomic language models are trained on genomic sequences created by evolution. The biological processes that led to these sequences are often not known to anyone, but an AI can still infer them from the sequences. Fundamentally, genomic language models are designed to discover knowledge we don’t yet have.
Progress hasn’t been just theoretical. Many groups are looking into how to engineer proteins and create new drugs with AlphaFold and protein language models. But at the same time, how we think about diseases and how we diagnose and treat medical conditions hasn’t changed that much, so we still have a way to go before we can say that AI has truly transformed biomedicine.
Getting biomedical AI out of the playground
A lot of AI work these days is essentially engineering. It boils down to training bigger and more powerful models – which has actually been a reliable way to improve the capabilities of general-purpose AI. But making things bigger isn’t enough when it comes to more specialized models in domains like biomedicine.
General AI has such a large user base that we can just hand people strong models and they will quickly figure out what to do with them. But the number of people interested in proteins who have the technical skills to work with specialized AI is tiny. So if the developers of these specialized models haven’t clearly demonstrated what they are actually useful for, most chances no one will.
Instead of focusing on useful applications, a lot of biomedical AI work is revolving around benchmarks that are interesting but not that useful in practice. When we compare models based on, say, how well they classify protein superfamilies, it provides some information on which algorithms absorb more knowledge and might perform better overall, but experimental biologists have no idea how to use this to discover new drug targets.
So it looks like we have made decent progress on training powerful AI models that absorb relevant biological knowledge, but not enough progress on figuring out how to use them. Our progress has also been somewhat restricted to proteins. There are many other areas in biology that could use AI.
One area that I think deserves more attention is human genetics. Human geneticists have been struggling with real, thorny problems for a long time, and the kind of challenges they are dealing with are exactly the kind that AI is really good at.
What’s been bothering human geneticists
Genetics is more difficult to study in humans than in other organisms. If geneticists want to know what makes some tomatoes more red and juicy than others, they can just breed, clone and genetically modify these tomatoes. At least in theory, it’s not difficult to come up with a clean experiment that would tell us exactly how specific genetic variants affect specific traits. But we can’t do direct genetic experiments on humans because they’re unethical. Human geneticists have to resort to indirect evidence from association studies, cell experiments and animal models, which leads to many conceptual and technical challenges.
It’s still possible to understand genetic effects in humans, but it takes a lot of hard work. Consider the genetics of obesity. The leptin hormone, which regulates fat cells and appetite, was first discovered in extremely obese mice. Follow-up experiments confirmed that damaging this gene makes mice obese, and association studies showed that people with similar mutations had the same phenotype (severe early-onset obesity). Many experiments later, a handful of other genes in the same pathway were also connected to obesity based on similar evidence. That’s as clear evidence as we can get for genetic effects in humans. But genome-wide association studies have found hundreds of other genomic regions correlated with obesity, and who has the time to breed mice and put the same amount of effort for each one of these suspicious regions? And who knows whether these genetic effects work the same way in humans and mice.
I’m not expecting animal experiments to become obsolete, but I imagine we can be much more efficient in how we use them. To draw an analogy from proteins: we sometimes still need crystallography to figure out protein structure, but in the age of AlphaFold we can often skip this and save time and money.
Why human genetics needs AI
As I tried to show with the obesity example, human genetics relies on evidence from many different sources. We need to consider evidence like genetic differences between people, evolutionary trends across species, and the results of experiments on animals and cells. It requires piecing together a lot of knowledge and data. And that’s something modern AI is quite good at.
That’s a somewhat philosophical reason to believe AI can make sense of genetic effects. Let’s see a few concrete examples of how this can work in practice, starting with evolutionary data.
As of 2024, we know of hundreds of millions of gene sequences from hundreds of thousands of species. Most of this huge data is just DNA and protein sequences, with no information on what each sequence actually does. It sounds like very superficial data that couldn’t teach us much about genetic effects. But it can.
By comparing how the same gene looks in different species, we can see which parts are conserved by evolution and don’t change much. Conserved regions are generally more important, meaning that mutations in these regions are more likely to cause disease. So just by looking at genomic sequences across species we can learn something important about genetic effects.
In a recent work of mine we took this idea a step further and showed that protein language models learn which genetic variants cause disease even in genes with limited evolutionary data. The AI learns what kind of sequences are and aren’t tolerated by evolution just by seeing millions of genes repeating in countless different forms across organisms. It’s a good example of how an AI can piece together huge amounts of seemingly shallow data to learn something useful that has clinical implications, like diagnosing genetic disorders.
If done right, AI can not only make genetic research more efficient, but also reduce some of the bias that feeds into it. An example of such bias is that almost 80% of participants in genetic studies are of European ancestry, despite making up only 16% of the world’s population. One of the reasons I like the use of protein language models in genetics is that they aren’t affected by this kind of bias. In fact, 99.99% of their training data comes from non-human organisms.
So we’ve seen how AI can take one type of data (genome sequences in different species) to learn about genetic effects. But there are many other types of relevant data. For example AlphaMissense was trained on genetic differences within the human population (in addition to sequences from different species) to predict whether mutations are damaging.
The data that is likely most relevant to genetic effects is the millions of individuals for which we now have both genetic information and medical history. Statistical genetics typically uses these records in isolation from other data, but it makes sense to combine it with other genomic data. AI can be a good vehicle to combine the genetic and genomic data. For example, we can train an AI on other types of genomic data to predict which kinds of mutations are more likely to have an effect, and then use it to find genetic effects with less direct data.
These examples demonstrate good progress we have made at finding genetics effects with AI and genomic data, but we are really only scratching the surface.
Making the match
I’m excited to see AI used in interesting ways in more areas of biomedicine. We have every reason to believe it can help with important problems like diagnosing genetic disorders, predicting risk for serious conditions, or coming up with new treatments.
If we are serious about this, the key part is to show exactly how AI can be used for each of these problems. We need to focus on important problems that biologists and clinicians truly care about and are struggling with, and spend less time chasing small performance improvements or finding out who has the biggest model.