Screening Of Biological Phase-separating Proteins With Machine Learning Models That Integrate Multimodal Features

Phase separation(PS)is one of the mechanisms that mediate the compartmentalization of biomacromolecules such as proteins and nucleic acids in cells,forming the biomacromolecular condensates or membraneless organelles(MLO).Therefore,the systematic identification of potential proteins that can undergo phase separation in cells is important for understanding the phase separation process and its biological mechanisms.In this study,we divided phaseseparating proteins into two groups according to the mechanism by which they undergo phase separation: PS-Self proteins can interact with the same protein species spontaneously to form droplets(this process is also known as protein self-assembly),while PS-Part proteins interact with partners to undergo phase separation.Existing PS predictors,when evaluated on two test protein sets,preferentially predicted self-assembling proteins.However,most phaseseparating systems demand multiple partners in biological conditions.Thus,a new predictor is required for screening potential PS-Part proteins.Herein,we compared the amino acids(AAs)composition of the two phase-separating protein groups and found that there were differences in sequence patterns between them.We propose that properties other than sequence composition can provide crucial information in screening phase-separating proteins.By incorporating phosphorylation frequencies and immunofluorescence(IF)image-based droplet forming propensity with other PS-related features,we built two independent machine-learning models to separately predict the two protein categories.Results of independent testing suggested the superiority of integrating multimodal features.Further validation on the proteome of MLOs confirmed the ability of our models to identify partner-dependent phase-separating proteins.We also performed experimental verification on the top-scored proteins DHX9,Ki-67 and NIFK.Their phase separation behavior in vitro revealed the effectiveness of our models in PS prediction.We implemented a web server named Pha Se Pred(http://predict.phasep.pro/)that incorporates our two models together with representative PS predictors.Currently,the website incorporates comprehensive scores of over 100,000 proteins from 18 species and provides quantiles for different PS-related features at the proteomic level.In conclusion,this study realizes the prediction of PS-Self and PS-Part proteins based on the phase separation mechanism and incorporates multimodal features to improve the accuracy of the PS prediction model.In addition,Pha Se Pred,a meta-predictor that integrates multiple PS predictors,can provide crucial information for profiling phase separation propensity and providing crucial information for the identification of candidate proteins.