Where AI-designed antibodies break down | AI Learning Guide

These failures share a common cause: the model is being asked to do something its training data did not prepare it for. The generalization gap is the clearest case. A model's prediction is grounded when the problem resembles its training data and becomes an extrapolation when it does not, but the output looks the same in both cases, so there is no built-in warning. Over-optimization is a different route to the same problem: if the scoring function rewards predicted affinity, the search will find sequences that satisfy the function rather than sequences that bind. Developability failure is the one people forget, because a strong predicted binder can still aggregate or express badly. And target-class transfer matters because a model trained on soluble proteins has little to say about a membrane protein.

0:00 / 0:00

Auto-play

The generalization gap, stated plainly

A model's prediction is trustworthy in proportion to how similar the problem is to its training data. For a target that resembles known antibodies' targets, the prediction is grounded. For a target unlike anything in the training set, the model still produces a confident-looking number, but that number is an extrapolation. The danger is that the output format does not distinguish the two cases, so a low-quality prediction looks the same as a high-quality one.

Recognizable failure patterns

Generalization gap: performance degrades on targets or epitopes unlike the training data, without an obvious warning in the output.
Over-optimization: the search exploits the scoring function, producing sequences that score well on the proxy but bind poorly in reality.
Developability failure: predicted binding is good, but the molecule aggregates, expresses poorly, or is unstable.
Target-class transfer: models trained on soluble antigens transfer poorly to membrane proteins or conformation-specific epitopes.