What the Loop Cannot Do Alone | AI Learning Guide

The loop improves with every round, but it cannot run itself. Start with the most basic point: the feedback path needs measurements, and measurements need experiments. No amount of model refinement substitutes for actually running the assay. Then there is interpretation. Suppose a candidate binds weakly — do you throw it out, or keep it because its scaffold is promising and worth optimizing? That call depends on chemical intuition and on what the project is trying to achieve, and a score cannot make it. Finally, the criteria are chosen by people. Deciding that stability now matters more than raw affinity is a statement about what kind of drug you want. So the model does one job at a scale no team could match, and the laboratory does a different job that no model can do. The loop is where those two jobs meet.

0:00 / 0:00

Auto-play

The pipeline as a cycle

Generate, score, filter, synthesize, measure, feed back. Each pass through the cycle narrows the uncertainty about which molecules are worth pursuing, and each pass starts from a model and a set of criteria that the previous round improved. The cycle ends not when the model is perfect but when a candidate is good enough to advance.

Where judgment stays human

Experiments must actually be run — the feedback path depends on real measurements, not on more computation
Interpreting a result is a decision: a weakly binding candidate may be discarded or kept for optimization depending on its scaffold and the project's goals
The criteria are chosen, not derived — deciding that stability now outweighs raw affinity reflects what kind of drug is wanted

A division of labour, not a replacement

The model's contribution is scale: it proposes and ranks candidates far faster than any team could screen by hand. The laboratory's contribution is truth: it establishes which predictions hold. Neither replaces the other, and the loop is where the two meet.