The first wave of AI structural biology answered a question chemists had chased for fifty years: given a sequence, what shape does the protein fold into? That was necessary, and it was not sufficient. Knowing a lock's shape does not tell you which key turns it. The interaction era is about the keys.
From Shape to Affinity
Structure prediction tells you what a target looks like. Drug discovery needs something harder: a reliable prediction of how strongly a candidate molecule binds, and what happens when it does. Models that estimate binding affinity and interaction directly — rather than inferring it from static structure — move the field from "here is the lock" to "here is which keys are worth synthesizing."
Why This Reshapes the Funnel
Traditional discovery is a brutal attrition process: synthesize many candidates, test them in the lab, and watch almost all of them fail. The expensive step is the wet lab. If an interaction model can rank candidates by predicted affinity before anyone synthesizes anything, the funnel inverts:
- Triage in silico first. Spend lab budget only on the candidates the model ranks plausible.
- Explore wider, cheaper. Computational screening can consider chemical space far larger than any synthesis program.
- Fail faster, fail cheaper. A bad candidate eliminated by a model costs compute, not months.
The Honest Caveat
Prediction is a filter, not an oracle. The wet lab remains the arbiter of truth, and confident-but-wrong predictions are a real failure mode. The value is not that the model is always right — it is that the model is right often enough to change where you spend the expensive resource. That is what "interaction era" actually means: not the end of the lab, but a far smarter decision about what to put into it.