Liquid-liquid phase separation (LLPS) of biomolecules has recently emerged as a crucial mechanism underpinning the formation of biomolecular condensates (or membrane-less organelles) for cellular organization. Dysregulation of biomolecular LLPS has been conceived  closely implicated in a number of disorders. With several databases about proteins related LLPS released timely, a numbe of tools for predicting phase separating proteins were developed. In this work, based on the data from newly updated database LLPSDB v2.0, we compare ten predictors of proteins undergoing LLPS, and highlight the advantages and limitations for the construction of them. The results indicate that the PSPredictor, FuzDrop and DeePhase, the new generation methods that were built upond deep learning techniques, always outperform other tools on different negative test datasets. Disorder regions play a crucial role in phase behavior of proteins. However, all the predictors could not predict LLPS propensity of proteins well when correlated with protein saturation concentrations. Further investigation, such as extracting valid features of sequence pattern or combining physicochemical properties of protein sequence with LLPS experimental conditions may improve the current prediction tools.

protein, liquid-liquid phase separation, predictor