Why a promising molecule fails in the lab or in animals | AI Learning Guide

Take the off-target case first, because it is the clearest. Many proteins have similar binding pockets, so a molecule designed for one can fit another. If that other protein matters — say, one involved in heart rhythm — you get a side effect, and sometimes a serious one. The model did not score that, because it was only looking at the intended target. Toxicity works the same way: it usually does not show up in binding data at all; it emerges from how the molecule behaves in living tissue. And pharmacokinetics is the quiet killer. A molecule that binds perfectly but is destroyed by the liver before it reaches the target is useless. None of these are model errors — they are failures the model was never asked to predict.

0:00 / 0:00

Auto-play

Common biological failure modes

Off-target binding — the molecule attaches to proteins other than the intended target, producing side effects
Toxicity — the molecule or its breakdown products damage cells, organs, or DNA
Poor pharmacokinetics — the molecule is broken down, cleared, or never reaches the target tissue
Rapid metabolic breakdown — the body converts the molecule into inactive or harmful compounds

Why the model does not catch these

Each failure mode depends on the whole body, not just the target. Off-target binding depends on which other proteins are present and how similar they are. Toxicity depends on how cells respond. Pharmacokinetics depends on absorption, distribution, metabolism, and excretion. The model can estimate some of these as properties, but the estimate is a guess about a living system, and the living system is what decides.