A hybrid model for multiscale prediction and generation of phase separating protein regions

Liquid-liquid phase separation (LLPS) is emerging as a fundamental process supporting multiple facets of biological systems. This phenomenon enables the dynamic compartmentalization of biomolecules contributing to a wide range of cellular functions, though in many instances its precise role and evolution remain unclear. Protein phase separation naturally occurs within cells and is prevalent across all species. Despite a recent surge in protein LLPS discovery, current predictive models lack generalizability and fail to identify the full spectrum of phase-separating proteins. To address this shortcoming, we developed Phaseek, a hybrid model integrating contextual sequence encoding with statistical graph representations to score LLPS propensity of amino acid sequences. Phaseek accurately identifies phase-separating proteins across diverse biological contexts, predicting key functional regions and the effects of point mutations. Proteome-wide predictions for 18 species highlight important physicochemical features. Gene Ontology enrichments recapitulate known processes (e.g., nucleic acid binding, nuclear localization, chromatin organization) and suggest novel areas of investigation. Phylogenetic analysis of orthologs further suggests that LLPS is evolutionarily conserved beyond sequence similarity. In addition, we used Phaseek to design de novo phase-separating peptides and achieved a 70% success rate in vivo. Provided with a user-friendly implementation, Phaseek serves as a multipurpose LLPS predictor for advancing both fundamental and applied LLPS research.