AI has made stunning breakthroughs in predicting protein structures and designing new proteins, but it still cannot predict how individual cells and tissues behave in the human body, according to leading computational biologists. This fundamental limitation means the field is likely decades away from using AI to truly transform medicine, despite widespread hype suggesting otherwise.

Andrea Califano, a preeminent computational biologist at Columbia University, recently published a perspective in the journal Cell calling on the scientific community to refocus AI research on 15 critical, biologically relevant problems. His critique challenges both the AI and biology communities to stop celebrating incremental wins and start tackling the harder questions that could actually change how we treat disease.

What's Wrong With Today's AI Biology Models?

The problem runs deeper than simply needing more data or computing power. Most AI models used in biology today are based on large language models, the same technology that powers ChatGPT and Claude. These models work by analyzing sequences of tokens arranged one after another, which works well for proteins (sequences of amino acids) and DNA (sequences of nucleic acids).

But here's where biology breaks the model's assumptions. In language, words near each other in a text are strongly related, while words far apart have little connection. Biology doesn't work that way. Two genes might regulate each other's function but sit far apart in the genome. As Califano explained, biology is more like "a bag of groceries where purchases are not organized in sequential fashion".

As Califano

"AI models in biology are based on large language models, the same models that power chatGPT, 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," explained Andrea Califano, the Clyde and Helen Wu Professor of Chemical and Systems Biology at Columbia University.

Andrea Califano, Clyde and Helen Wu Professor of Chemical and Systems Biology at Columbia University

The second major problem is how researchers test their AI models. Instead of asking whether AI can solve critical biological mysteries, scientists often demonstrate their models by showing they can distinguish between different types of immune cells. But antibodies already do that job well. Testing on easier, less relevant problems masks the real limitations of the technology.

Can We Just Feed AI More Data to Fix This?

The short answer is no. The search space of possible biological combinations is so vast that no amount of additional data will solve the problem. When you consider all possible combinations of genes, RNA molecules, proteins, and metabolites, the number explodes to a size vastly larger than the number of atoms in the universe. There simply isn't enough data to train a model that explores every possibility.

This is why Califano argues the AI community needs to take a different approach. Rather than building general-purpose models that try to solve any biological problem, researchers need to "teach biology to the AI" by pre-wiring models with biological knowledge. This means encoding what we already know about how cells work into the AI system itself, rather than expecting the model to discover it from raw data.

How to Refocus AI Research on Biology's Biggest Problems

• Embed Biological Knowledge: Instead of training models on raw data alone, incorporate existing biological principles and cellular mechanisms directly into the AI architecture so it doesn't waste computational resources exploring impossible scenarios.
• Test on Hard Problems: Move beyond simple classification tasks and evaluate AI models on whether they can predict cell behavior, identify disease drivers, or suggest which patients will respond to specific drugs, not just whether they can label cell types.
• Focus on Specific Solutions: Rather than pursuing a universal "virtual cell" that solves all biological questions, prioritize AI systems designed to solve one critical problem extremely well, such as predicting cancer drug response or identifying disease mechanisms.

Califano's call to action draws inspiration from David Hilbert's famous 1900 list of 23 mathematical problems that shaped mathematical research throughout the 20th century. By identifying 15 critical biological questions and rallying the community around them, Califano hopes to redirect AI research toward problems that could genuinely transform medicine.

Why Does This Matter Right Now?

The tension between what biologists need and what AI researchers want to build is becoming urgent. Current drug discovery and development relies on making hypotheses, testing them in animal models, and hoping human biology matches. This approach worked well for finding obvious disease drivers, like EGFR mutations in lung cancer. But most cancers don't have such clear culprits. The patterns that drive disease are incredibly complex, and our current methods struggle to pinpoint them.

Meanwhile, every failed clinical trial and every missed opportunity to predict which patients will respond to a drug represents human suffering that better predictive models could prevent. Biologists want AI systems that are highly specific and can solve one critical problem extremely well. The AI community, by contrast, wants to build generalizable models that can tackle any problem in biology without focusing on anything specific. A universal "virtual cell" may take decades to develop, but patients need better treatments today.

"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," stated Andrea Califano.

Andrea Califano, Clyde and Helen Wu Professor of Chemical and Systems Biology at Columbia University

The message from one of the world's leading computational biologists is clear: AI's potential in biology is real, but the path forward requires fundamental changes in how models are designed, tested, and deployed. The hype may need to cool, but the opportunity to transform medicine through smarter AI remains within reach, if the field is willing to do the harder work of teaching AI actual biology.