Predicting Structure and Binding | AI Learning Guide

The key distinction on this page is that structure prediction and binding prediction are not the same job. Structure prediction tells you the shape a sequence would fold into, and shape matters because binding is geometric: the two surfaces have to fit. But fitting is not the same as binding well. Binding prediction tries to answer the harder question of whether this candidate actually binds this target, and it is trained on pairs with known outcomes. Now the practical point. Both models were trained under specific conditions, and they are reliable when your case resembles those conditions. Push outside that range and the estimate degrades quietly. So a prediction is a hypothesis you still have to test, not a result you can trust on its own.

0:00 / 0:00

Auto-play

Two questions, two models

Structure prediction answers "what shape would this sequence take, and does its binding site complement the target?" Binding prediction answers "would this candidate actually bind, and how well?" They are separate because geometry and affinity are separate facts: a well-folded binding site can still fail to bind a given target, and a model that predicts shape well is not automatically good at predicting affinity.

Where predictions get shaky

A prediction model is only as good as the match between its training conditions and the case in front of it. Unusual targets, unusual antibody families, or assay formats the model never saw push the estimate outside the range where it was validated. Treat a prediction as a hypothesis to test, not as a result.