This work was led by @jacoberts.bsky.social in collaboration with @ben-eysenbach.bsky.social and @crji.bsky.social. It would not have been possible without funding from the United States federal government, via the NIH, NSF, DOE, and AFOSR.

This method is now built into Foldy, our lab's open-source protein engineering platform. Other updates: Foldy uses Boltz-2x for structure prediction, runs ESM family models, and is deployable with a single command. Setup instructions: github.com/JBEI/foldy

Where did the improvements come from? We show that the biggest factor is a new policy called naturalness warm-start, a way to pretrain the activity predictions with the outputs of the ESM family of protein language models.

We present FolDE, an ALDE method designed to maximize end-of-campaign success. Across 20 ProteinGym datasets, FolDE discovers 23% more top 10% mutants than the random forest-based ALDE baseline (p=0.005) and is 55% more likely to find top 1% mutants.

We've observed that existing zero-shot and few-shot protein activity prediction methods often select batches of very similar mutants. We found that selecting closely related mutations narrows the data used to train subsequent models, thereby weakening predictions in later rounds.