What I mean when I say “AI for biology”
A small taxonomy, because prediction tools, design tools, and lab automation are solving very different problems.
“AI for biology” is used to describe everything from a protein structure predictor to a robot that pipettes. The ambiguity is not harmless. It makes it easy to compare unlike things, and it makes it hard to ask what a particular system is actually for.
Here is the taxonomy I find useful.
1. Measurement and representation
These systems help us see. They denoise images, embed cells, annotate sequences, and align modalities. Their output is a representation that a scientist can use, not a decision. The relevant evaluation question is whether the representation preserves the distinctions that matter downstream.
2. Prediction
These systems estimate a property: structure, binding affinity, expression response, toxicity. Prediction is the most crowded category and the one most often confused with translation. A prediction is only as useful as the decision it informs, and the decision is usually made under uncertainty that the prediction does not capture.
3. Design
These systems propose interventions: a sequence, a molecule, a perturbation. Design changes the burden of proof. A predictor can be wrong and still useful for triage. A design tool that proposes something expensive to make needs to be right often enough to justify the experiments it triggers.
4. Automation and control
These systems run the lab: scheduling, liquid handling, adaptive experiment design. Their value is throughput and consistency, and their failure modes are operational rather than statistical. A model that is slightly worse but reliable every day can be far more valuable than a better model that is fragile.
Why the distinction matters
The categories have different bottlenecks. Measurement is limited by data quality. Prediction is limited by distribution shift. Design is limited by validation cost. Automation is limited by integration and trust.
They also have different defensibility. A representation model is easy to copy. A closed-loop lab that generates its own data is not. A design tool is only as good as the experimental pipeline that tests its output.
When someone says “we use AI to discover drugs,” it is worth asking which of these four things they actually do. The answer usually clarifies both the science and the business.
The phrase is fine as a shorthand. It is a poor substitute for saying what the system measures, what it predicts, and what happens when it is wrong.