What a Scoring Model Actually Predicts | AI Learning Guide

Think of the scoring model as a very well-read predictor rather than an instrument. It has seen thousands of molecule-and-target pairs where the binding was already measured in a lab, and from those examples it learns to map a structure onto an estimated property. The headline property is binding affinity, how tightly the molecule sits in the target's pocket. But notice what it never does: it never touches the protein. It reads a structure and returns a number. That number is trustworthy only when the candidate resembles the chemistry the model was trained on. Feed it something unusual and it will still answer confidently, because a model has no way to say I don't know.

0:00 / 0:00

Auto-play

A scoring model predicts a property of a candidate molecule — most importantly binding affinity — by learning from known molecule–target pairs. It estimates; it does not measure.

What goes in, what comes out

The input is a representation of the candidate, typically the graph or string form from the previous chapter. The output is a number, or a small set of numbers, standing for properties the model was trained to predict. Binding affinity is the headline quantity because it is the property the whole pipeline exists to improve.

A prediction is only as good as the training data behind it. When a candidate falls outside the chemistry the model has seen, the model does not report ignorance — it returns a number anyway. Treating that number as reliable is one of the standard ways a computational shortlist goes wrong.