It’s Time to Teach AI Some Biology

Columbia’s Andrea Califano, one of the world’s preeminent computational biologists, has no doubt that AI will eventually help biologists make tremendous discoveries.

But unlike many AI enthusiasts, Califano, who is known for developing mathematical models that identify key drivers of cancer, thinks that future is probably decades away, not right around the corner.

As he sees it, AI models in biology need a fundamental readjustment to succeed and the AI community needs to begin testing their models on more biologically important questions.

Califano just published a perspective on AI in biology, appearing in the journal Cell, challenging the scientific community to focus on addressing 15 critical, biologically relevant problems. His list draws inspiration from David Hilbert’s 1900 list of 23 mathematical problems that shaped mathematical research throughout the 20th century.

We recently spoke to Califano, the Clyde and Helen Wu Professor of Chemical and Systems Biology at Columbia University Vagelos College of Physicians and Surgeons and head of Biohub New York, about his call to action.

What prompted you to write this critique of AI in biology?

Well, a number of things. I think AI in biology has been really spectacular in many ways: it has advanced our ability to predict protein folding and structure and design completely new proteins. But there is also a tremendous amount of hype about AI right now.  

In medical research, what we really want to do is predict the behavior of individual cells and tissues and control or change those behaviors to improve human health. It’s a largely unmet goal.

AI has the potential to accelerate discovery, but the AI models that dominate biology today haven’t shown any ability to predict the behavior of cells, despite the sometimes wild statements made about them.

You know, I have no doubts that AI will do great in biology—10 years from now, 20 years from now. I wrote this to generate discussion and help the community focus their attention on a handful of problems that, if solved, could change the face of biology and medicine.

What changes are needed to make AI a success in biology?

AI models in biology are based on large language models, the same models that power chatGTP, Claude, and other popular chatbots. And the word “language” tells you everything. Language is something where you have a string of tokens sequentially organized, one after the other. Proteins are exactly like that; they are sequences of amino acids. DNA is a sequence of nucleic acids.

LLMs work really well in those cases because they use locality in their “attention model” to make predictions. That is, tokens are most strongly related to nearby tokens. I’ll give you an example: A word in the middle of a book is likely strongly associated to other words next to it but not so much to a word in the final page of the book.

In the context of cell behavior and function, this concept of locality does not really apply. You may have two gene products that functionally interact—for instance, one may regulate the other—but are physically very far apart from each other in the genome. Biology is more like a bag of groceries where purchases are not organized in sequential fashion.

The other issue, in my opinion, is the way people assess the capabilities of their AI methodologies. Rather than testing them on whether they can make critical biological discoveries they try to show how well they can differentiate a CD4 from a CD8 T cell. We don’t need AI to do that; antibodies already do a great job. It’s OK to use that as an initial benchmark, but we should start introducing harder and more biologically relevant tests.

Can we just feed more data into the models and let them figure it out?

Well, the problem is that with the current models there will never be enough data because the search space it has to explore is much vaster than the number of atoms in the universe. It’s just not possible, and that’s the problem we need to figure out.

We need to generate the models that will actually work. I think we need to sit down and start doing a little bit more thinking.

I think if we want AI to work in biology, we need to teach biology to the AI, essentially pre-wiring it with biological knowledge. Otherwise, your model would try to look at all possible combinations of genes, RNAs, proteins, and metabolites, and that number explodes very quickly to a size where there is simply not enough data to train the model.

Why is it so important to get AI on the right path?

The way we do research now—we make hypotheses, we test them, often in animals that don’t always share human biology—is becoming increasingly ineffective. It was good for discovering the low-hanging fruits. Finding that EGFR mutations drive about 15% of lung cancers was relatively easy, we only had to sequence genomes and do some simple statistics. Unfortunately, most lung cancers do not have EGFR as a driver mutation and the pattern that induces tumorigenesis is incredibly complex and difficult to pinpoint.

But there’s a tension that exists in the community right now:  The biologist wants models that are highly specific and can solve one critical problem extremely well, whereas the AI community wants to build generalizable model that can solve any problem in biology without focusing on anything specific.

A universal “virtual cell” may take decades. In the meantime, we need to cure cancer today. Every failed clinical trial, every failure to predict patient response or select the best drug, represents human suffering that better predictive models could alleviate.