What AI can and cannot do for drug design

A few years ago, predicting the shape of a protein from its sequence was one of the great unsolved problems in biology. Then AlphaFold arrived, and the problem was, for many practical purposes, solved. The next step came quickly: if a model can predict structure, can it design a protein that does not exist yet, built to bind a chosen target?

It can. That question is at the centre of my work as Head of Computational Biology and AI at SEVO Bioscience, where we partner with Mila, the Quebec AI Institute, on AI-designed therapeutic proteins. I will not discuss the specifics of that work here. But I can share how I think about what these tools change, and what they do not.

What AI changes

The first thing AI changes is the size of the search. A human designer might test dozens of variants. A model can propose thousands, rank them and suggest which few deserve a place on the bench. That turns design from a slow craft into something closer to a search problem.

The second is speed. Ideas that once took a year to reach the lab can now get there in weeks. Faster cycles mean more chances to learn from failure, and failure is still most of the job.

What AI does not change

A design is a hypothesis. The cell is the test.

A protein that looks perfect on screen still has to be made, folded and purified. It has to reach the right tissue, survive long enough to act, and not provoke the immune system. None of that is guaranteed by a good score. Models are trained on what we already know, and medicine lives in the places where our knowledge is thin.

This is where my background in cell biology and drug delivery matters. A design is a hypothesis. The cell is the test. Many of the hardest questions, like whether the body will tolerate a new molecule, are the same ones I study with drug carriers.

Data is the real bottleneck

The best models are limited by the data they learn from. Public datasets are biased toward proteins that were easy to study. The groups that will do best are those that close the loop: design, test in the lab, feed the results back into the model, repeat. That loop needs biologists and machine-learning researchers in the same room, speaking each other’s language.

Hype and caution

Every few months, someone announces that AI has made drug discovery solved. It has not. It has made the early stages faster and wider. The clinic remains slow, expensive and humbling, as it should be when people’s health is at stake.

I am optimistic. I have spent my career at the bench and most of my life writing code, and this is the first time the two feel like one discipline. The models will keep improving. Our job is to ask them better questions and test their answers honestly.