🦖

Sampling diversity in protein folding models with evolutionary strategies

TLDR: Protein structure prediction models estimate a highly accurate energy function, but the optimization task is still hard — models can often come up with far more confident predictions when sampled on the order of 1000s of runs with dropout enabled. I propose using evolutionary strategies to find an optimal set of neurons to dropout, maximizing the confidence of predicted structures.

af-evolve.pdf494.4KB