Why a High Score Is Not Proof | Visual AI Learning Guide

Here is the uncomfortable part. Every score we have discussed is a prediction. When you plot predicted binding against what was actually measured in the lab, the points do not sit on a clean line — they spread out. Some compounds the model ranked highly turn out to do nothing. The spread is worst for molecules unlike anything in the training data, which is exactly the novel molecules generation produces. So a high score does not mean the molecule works. It means the molecule is worth the cost of an experiment. The score is a hypothesis, and the assay is what turns it into knowledge.

0:00 / 0:00

Auto-play

Both screening and generation depend on a scoring function, and a score is a prediction, not a measurement. When predicted binding is plotted against measured binding for compounds that have actually been tested, the points scatter widely around the ideal line. Some compounds the model loves turn out to be inactive; some it dismisses turn out to bind well.

The uncertainty has a specific cause. A model is reliable where it has seen many similar examples during training, and unreliable where it has not. Truly novel molecules — exactly what generative design produces — sit in the sparse region, so their scores are the least trustworthy precisely when they matter most. A high predicted score therefore means "this is worth testing," not "this works." The only way to convert a prediction into knowledge is to run the experiment.