PSTP: Decoding Latent Sequence Grammar for Protein Phase Separation through Transfer Learning and Attention

Abstract

Phase separation (PS) is essential in various biological processes, necessitating high-accuracy predictive algorithms to study numerous uncharacterized sequences and accelerate experimental validation. However, many recent prediction methods face challenges in generalizability due to their reliance on engineered features, and accurately identifying protein regions involved in PS remains difficult. We propose PSTP, which employs a dual-language model embedding strategy and a lightweight attention model. PSTP achieves identification of 84% of PS regions in PhaSePro and demonstrates ~50% improvement in correlation coefficient over existing models. It shows robust performance across different types of PS proteins, and shows the potential for predicting artificial proteins. By analyzing 160,000 variants, PSTP characterizes the link between the incidence of pathogenic variants and residue-level PS propensities. PSTP's predictive power and broad applicability make it a valuable tool for accelerating our understanding of biomolecular condensates, and their mechanisms underlying disease development.