Pseudo-perplexity in One Fell Swoop for Protein Fitness Estimation

TL;DR

This paper introduces a fast, single-pass method to estimate protein sequence fitness using language model embeddings, achieving near state-of-the-art accuracy and enabling efficient exploration of functional sequences.

Contribution

The authors propose the One Fell Swoop approach for pseudo-perplexity estimation, significantly improving computational efficiency and performance in protein fitness prediction tasks.

Findings

OFS pseudo-perplexity performs nearly as well as true pseudo-perplexity.

Achieves new state-of-the-art on ProteinGym Indels benchmark.

Effectively detects increased stability in ancestral protein sequences.

Abstract

Protein language models trained on the masked language modeling objective learn to predict the identity of hidden amino acid residues within a sequence using the remaining observable sequence as context. They do so by embedding the residues into a high dimensional space that encapsulates the relevant contextual cues. These embedding vectors serve as an informative context-sensitive representation that not only aids with the defined training objective, but can also be used for other tasks by downstream models. We propose a scheme to use the embeddings of an unmasked sequence to estimate the corresponding masked probability vectors for all the positions in a single forward pass through the language model. This One Fell Swoop (OFS) approach allows us to efficiently estimate the pseudo-perplexity of the sequence, a measure of the model's uncertainty in its predictions, that can also serve as a fitness estimate. We find that ESM2 OFS pseudo-perplexity performs nearly as well as the true pseudo-perplexity at fitness estimation, and more notably it defines a new state of the art on the ProteinGym Indels benchmark. The strong performance of the fitness measure prompted us to investigate if it could be used to detect the elevated stability reported in reconstructed ancestral sequences. We find that this measure ranks ancestral reconstructions as more fit than extant sequences. Finally, we show that the computational efficiency of the technique allows for the use of Monte Carlo methods that can rapidly explore functional sequence space.