Generalizable prediction of liquid-liquid phase separation from protein sequence

Liquid-liquid phase separation (LLPS) is a physicochemical process through which a homogeneous liquid solution spontaneously separates into distinct liquid phases with different compositions and properties. Driven by weak multivalent interactions, LLPS in living systems enables dynamic compartmentalization of biomolecules to promote or regulate various cellular processes. Despite recent advances, predicting phase-separating proteins and their key LLPS-driving regions remains limited by the versatility of models. Here, we developed Phaseek, a generalizable LLPS predicting model that combines contextual sequence encoding with statistical protein graph representations. Phaseek accurately predicts LLPS-prone proteins in diverse biological contexts and identifies key regions and effects of point mutations. Proteome-wide analysis across 18 species suggests evolutionary conservation of LLPS among orthologs and associated biological processes. Additionally, predictions by Phaseek highlight the key physicochemical and structural properties associated with LLPS. Provided as an open-access model with a user-friendly implementation, Phaseek serves as a multipurpose LLPS predictor for advancing fundamental and applied research.