A Systematic Evaluation of Protein Phase Separation Predictors across Diverse Protein Landscapes

Background: Liquid–liquid phase separation (LLPS) plays a central role in cellular regulation, with its dysregulation linked to numerous diseases. LLPS is also increasingly implicated in various biological contexts, such as virus replication. These findings have driven the development of numerous computational predictors to screen and identify phase-separating proteins from sequence and/or structural models. Despite the need for these tools, their performance across diverse biological contexts remains incompletely understood, complicating tool selection and result interpretation. Results: We performed a systematic comparative analysis of 9 LLPS prediction algorithms using multiple curated datasets comprising both LLPS-positive (LLPS+) and LLPS-negative (LLPS−) proteins. The datasets span multiple biologically relevant scenarios, including intrinsically disordered proteins, folded proteins, proteins with LLPS-abolishing variants, benchmark datasets, and viral proteins. We observed substantial variability in predictive performance across algorithms when assessing proteins of different structural classes such as LLPS+ and LLPS− folded proteins, LLPS-abolishing mutations, and viral proteins. Conclusions: These results demonstrate that LLPS predictor performance is strongly context dependent, leading to different predictors being optimal for different biological questions. For overall protein assessment, DeePhase and MolPhase provided the most consistently accurate predictions, being the least impacted by structural bias. For assessing the impact of small mutations on LLPS propensity, PSPHunter, a built-for-purpose algorithm, reliably predicts mutation impacts, with structure-informed algorithms PSPire and PICNIC also providing strong insight. Across all evaluated datasets, the findings highlight the need for well-benchmarked training and testing data that encompass a broad and diverse range of protein classes.