amyloid-predict and LLPS-predict: Predicting phase separation propensities in the intrinsically disordered proteome

Abstract

Amyloid formation and liquid–liquid phase separation (LLPS) are two important phenomena in cellular biology, linked to both normal physiological functions and various pathologies. Here, we present a computational framework that scores amyloid propensities (amyloid-predict) or LLPS propensities (LLPS-predict) from protein language model embeddings, enabling rapid proteome-wide annotation of peptides and residues. amyloid-predict achieves classification performance that exceeds existing AI and physics-based tools on a hexapeptide benchmark while enabling substantially faster high-throughput screening; notably, amyloid-predict is sensitive to subtle mutational effects and is influenced by sequence patterning and context rather than amino acid composition alone. We apply these protein language model classifiers to all the IDRs in the human proteome and uncover several protein categories with significant enhancement in amyloid and/or LLPS propensity, suggesting insights into the biological roles of these protein categories. For example, signaling receptors, carbohydrate-binding proteins, and Ca 2+ binding proteins are enriched in aggregation propensity, while mRNA-binding proteins, ribonucleoprotein complex, and nuclear matrix proteins are enriched in LLPS propensity. Interestingly, we observe patterns of both high amyloid and LLPS propensity in several amyloid-forming and prionic proteins. Together, these results provide side-by-side landscapes of LLPS and amyloid potential across the disordered human proteome while offering a rapid screening tool for basic biology, disease-mechanism studies, and rational design of peptide therapeutics.