Prediction of liquid-liquid phase separation proteins based on protein language model

Liquid-liquid phase separation (LLPS) enables biomolecules to form membraneless condensates critical for cellular functions like RNA metabolism and protein synthesis. Identifying LLPS-associated proteins is essential for understanding their roles in cellular organization and disease. Current prediction methods are hindered by complex and inefficient feature extraction processes. Here, we present a novel framework combining a protein language model ProtT5 with a KmerConv module for local sequence pattern detection and a multi-head attention mechanism for global sequence information extraction. This integrated approach achieves high predictive performance across multiple diverse datasets and generalizes effectively across different species. Our method provides a robust and efficient tool for systematic LLPS protein identification, advancing research into biomolecular aggregation and its implications for health and disease.