Catching Safety Problems Before the Lab | AI Learning Guide

The key idea is that binding the target well does not make a molecule safe. The body contains thousands of proteins, and some resemble the target closely. A candidate that also binds one of those changes something it was never meant to touch, and that is a common source of side effects. AI can estimate which other proteins a new candidate is likely to bind, by learning from molecules whose interactions have already been measured. Candidates that look likely to hit several unintended proteins can be dropped or redesigned before laboratory work starts. But notice the limit: this is an early warning, not a clearance. A molecule can bind only its target and still be broken down into something harmful, or cleared from the body too fast. Those risks are harder to predict, and a model can only flag problems it has seen before.

0:00 / 0:00

Auto-play

Why binding well is not enough

The target is one protein among thousands. A candidate that also binds a similar protein will change something unintended, and that unintended change is what shows up as a side effect. Predicting which other proteins a candidate is likely to touch is therefore a safety question, not a bonus feature.

What the prediction is worth

AI's safety predictions are early warnings, not clearances. They let the most obviously problematic candidates be dropped cheaply, but they cannot establish that a molecule is safe. That still has to be shown by testing, first in the laboratory and eventually in people.

A model can only flag risks it has seen patterns for. A genuinely new kind of problem may not appear in any prediction, which is one reason safety testing cannot be replaced by a model's output.