New theory proves convergence of AI drug discovery loop

Fackeldey and Schütte have published a landmark theoretical paper on closed-loop generative selection, the algorithm powering modern AI-driven drug discovery. The loop works by: a generative model proposes molecules, a fitness oracle scores them, the best are kept (elitism), and the model is retrained on this elite set. Despite widespread use, no rigorous convergence theory existed because retraining breaks the Markov property. The authors resolve this by enlarging the state space to recover a Markov structure, then prove almost-sure convergence and derive a runtime bound that decomposes the search into time spent escaping each fitness level. They also analyze the role of memory—how much past data is retained for retraining. When learning improves steadily with more data, deeper memory never hurts; when it does not, an exit-time analysis pinpoints the optimal memory depth and shows that excess memory can actually slow convergence—a surprising and practically important result.

The theory extends naturally to multi-objective search and noisy oracles, which are common in real drug discovery where experimental measurements are noisy. For light-tailed noise, they quantify how many repeated evaluations certify progress; for heavy tails, robust estimators restore guarantees. Crucially, the analysis recasts the problem in terms of oracle evaluations—the true bottleneck in drug design—and yields a concrete evaluation-minimal strategy. A reproducible study confirms the predictions, including the counterintuitive cost of too much memory. This work closes a major gap between practice and theory, giving practitioners rigorous guidance on how to configure their generative loops for maximum efficiency.