Repurposing Protein Language Models for Latent Flow-Based Fitness Optimization

TL;DR

This paper introduces CHASE, a novel framework that leverages pretrained protein language models and flow-based generative techniques to efficiently generate high-fitness protein variants, outperforming existing methods on key benchmarks.

Contribution

CHASE repurposes pretrained protein language models by compressing embeddings into a latent space and uses flow-matching for direct high-fitness variant generation, avoiding costly gradient-based sampling.

Findings

Achieves state-of-the-art results on AAV and GFP benchmarks.

Enables direct generation of high-fitness variants without predictor guidance.

Synthetic data bootstrapping improves performance in data-scarce scenarios.

Abstract

Protein fitness optimization is challenged by a vast combinatorial landscape where high-fitness variants are extremely sparse. Many current methods either underperform or require computationally expensive gradient-based sampling. We present CHASE, a framework that repurposes the evolutionary knowledge of pretrained protein language models by compressing their embeddings into a compact latent space. By training a conditional flow-matching model with classifier-free guidance, we enable the direct generation of high-fitness variants without predictor-based guidance during the ODE sampling steps. CHASE achieves state-of-the-art performance on AAV and GFP protein design benchmarks. Finally, we show that bootstrapping with synthetic data can further enhance performance in data-constrained settings.