BioUnfold #4 — Why AI Needs Biology Literacy

Biology-aware AI

You should find the right problem — not just the right solution.

AI has transformed many industries, but its applications remain context-specific. In drug discovery, it does not solve a math problem in isolation — it intervenes in a biological system.

Broadly, there are two ways AI can contribute to discovery:

  1. Helping discovery through analysis — making sense of experimental results.
  2. Helping discovery by skipping a step — running simulations or predictions that reshape experiments.

Both roles shape how decisions are made across the discovery pipeline — either by deepening what we learn from experiments or by changing which experiments happen at all.

Why analysis needs biological understanding

When analyzing experimental data, knowing what you are looking at and why it matters lets you sharpen success criteria and reduce false positives.

Identifying a ‘cluster’ means nothing on its own. Biology gives structure to the signal — telling us which patterns matter.

A small shift in framing can dramatically improve outcomes. Imagine a high-content screening (HCS) campaign: adding just a few extra compound controls can reveal hits acting through unexpected mechanisms of action. These insights only emerge through conversations between modeling and experimental teams.

In one screen, the analysis looked dull. My biologist colleague insisted on seeing it — and spotted patterns I had missed. That follow-up turned out to be valuable for the project. It was not a better algorithm that made the difference, but the conversation around the data.

Why simulation needs biological understanding

When AI is used to skip experiments, biological literacy matters even more. Every model carries hidden assumptions. Without understanding the biology, it is easy to optimize for the wrong thing.

This is like using a proxy reward function in reinforcement learning. It can work, and often does. But the more innovative the project, the more those shortcuts limit what is possible.

A model trained without biological grounding may optimize beautifully for the wrong objective.
A model built with biological literacy can redefine the objective itself.

AI without biology can make experiments cheaper.
AI with biology can make experiments smarter.

You do not need to be a modeling expert or a bench scientist to start bridging these gaps — you just need the right conversation at the right time.

A conversation starter

Three questions I keep coming back to: