Why one molecule is never the answer | AI Learning Guide

The key idea here is that generation is necessary but not sufficient, and the reason is built into what the model knows. It learned which structures resemble real chemistry. It never learned which structures bind a particular target, survive in the body, or can be synthesised — those facts were not in its training signal. So a generated molecule can be perfectly plausible and still be a dead end. That is why the output is a pool rather than an answer, and why diversity inside the pool matters: if every candidate is a variation on one scaffold, the scoring stage has nothing to choose between, and you are back to a single guess.

0:00 / 0:00

Auto-play

The output is a pool, not an answer

Because generation is sampling from a broad distribution, the deliverable of this stage is a large, varied set of candidate structures. Its value lies in the range it covers, not in any single member. Narrowing that set to something worth testing is a separate job, done by scoring and filtering.

Necessary, but not sufficient

A generative model is necessary because a fixed library cannot contain a molecule nobody has made yet, and the search has to reach beyond what exists. It is not sufficient because plausibility under the model's distribution is not the same as being a good drug. The model has no access to binding measurements, safety data, or synthesis routes; it only knows which structures resemble the chemistry it was trained on. High-quality generation is a starting condition for design, not a result.

Diversity is a working requirement, not a nicety

A pool of near-identical molecules collapses the search back to one guess. Diversity is what makes the next stage meaningful: scoring can only rank candidates apart if the candidates actually differ in the ways that matter.