Why the Score Never Decides Alone | AI Learning Guide

The example is the clearest way to see the division of labour. A model trained on small-molecule pills has almost nothing in its history that resembles an injected biologic, so its score for that candidate is an extrapolation rather than a real estimate. The team knows this, and that knowledge is exactly what the model cannot supply. That is why the score informs the decision instead of dictating it — and why treating it as an automatic veto would ask the model to do something it was never validated to do.

0:00 / 0:00

Auto-play

What each side can see

The model sees measured properties and historical outcomes. The team sees mechanism, prior programs, competitive context, and the specific biology of the target. Neither view contains the other, so combining them is not a courtesy — it is the only way to cover the full evidence.

Disagreement is a signal, not a tie to break

When a prediction conflicts with expert reading, the useful move is to ask why. A compound flagged as high risk that experts believe in may share a surface property with past failures while differing in the mechanism that actually mattered. Resolving that question often improves the team's understanding of the compound, regardless of which side turns out to be right.

When judgment should override the score

A model trained mostly on small-molecule oral drugs has little to say about a biologic delivered by injection, because almost nothing in its training history resembles that candidate. Its score there is an extrapolation, not an estimate. A team that knows the candidate sits outside the model's training territory should treat the score as weak evidence and let the experimental data carry the decision.

Treating a prediction as a veto — dropping any compound above a score cutoff without review — converts a useful signal into a rule the model was never validated to support.