Machine learning models for predicting protein condensate formation from sequence determinants and embeddings

Abstract

Intracellular phase separation of proteins into biomolecular condensates is increasingly recognised as an important phenomenon for cellular compartmentalisation and regulation of biological function. Different hypotheses about the parameters that determine the tendency of proteins to form condensates have been proposed with some of them probed experimentally through the use of constructs generated by sequence alterations. To broaden the scope of these observations, here, we established an in silico strategy for understanding on a global level the associations between protein sequence and condensate formation, and used this information to construct machine learning classifiers for predicting liquid-liquid phase separation (LLPS) from protein sequence. Our analysis highlighted that LLPS-prone sequences are more disordered, hydrophobic and of lower Shannon entropy than sequences in the Protein Data Bank or the Swiss-Prot database, and have their disordered regions enriched in polar, aromatic and charged residues. Using these determining features together with neural network based word2vec sequence embeddings, we developed machine learning classifiers for predicting protein condensate formation. Our model, trained to distinguish LLPS-prone sequences from structured proteins, achieved high accuracy (93%; 25-fold cross-validation) and identified condensate forming sequences from external independent test data at 97% sensitivity. Moreover, in combination with a classifier that had developed a nuanced insight into the features governing protein phase behaviour by learning to distinguish between sequences of varying LLPS propensity, the sensitivity was supplemented with high specificity (approximated ROC-AUC of 0.85). These results provide a platform rooted in molecular principles for understanding protein phase behaviour. The predictor is accessible from https://deephase.ch.cam.ac.uk/ .

In silico prediction of in vitro protein liquid–liquid phase separation experiments outcomes with multi-head neural attention

“…PSPer was based on empirical rules defined for a small class of proteins (FUS-like family) (Wang et al, 2018) because very little data about other LLPS proteins were available at the time. Other sequence-based LLPS predictors have also been developed afterwards, using Machine Learning (ML) classifiers such as Random Forest (Saar et al, 2020) or Support Vector Machines (Sun et al, 2019). A deeper analysis of the current state of the art LLPS predictors from sequence only is available in Vernon and Forman-Kay (2019) Recently, more data related to LLPS proteins have been gathered in publicly available databases (Mészáros et al, 2020;You et al, 2020).…”

“…To the best of our knowledge, Droppler is the first model designed to predict the likelihood of proteins to undergo LLPS given a specific set of experimental conditions because existing approaches (Saar et al, 2020;Sun et al, 2019;Vernon and Forman-Kay, 2019;Orlando et al, 2019) focus only on the protein sequence and predict some sort of average LLPS propensity or in very specific experimental conditions. For instance, in Saar et al (2020) the authors only predict the LLPS propensity of a protein in "nearly physiological conditions".…”

“…To the best of our knowledge, Droppler is the first model designed to predict the likelihood of proteins to undergo LLPS given a specific set of experimental conditions because existing approaches (Saar et al, 2020;Sun et al, 2019;Vernon and Forman-Kay, 2019;Orlando et al, 2019) focus only on the protein sequence and predict some sort of average LLPS propensity or in very specific experimental conditions. For instance, in Saar et al (2020) the authors only predict the LLPS propensity of a protein in "nearly physiological conditions". Thanks to the explicit modeling of the experimental conditions, Droppler could be used instead to directly prioritize the in-vitro experiments that should be performed for i) the discovery of novel LLPS proteins and ii) to efficiently explore the experimental conditions space in order to find the physico-chemical settings that maximize the likelihood of LLPS for each specific protein sequence.…”

BIAPSS - BioInformatic Analysis of liquid-liquid Phase-Separating protein Sequences

“…This is because many different combinations of relevant interactions seem to be contributing to phase separation without anyone being universally necessary [12]. So far, however, with a few exceptions [13][14][15][16] mostly case-by-case studies of different sequences are performed, with the broader context of many findings, including their statistical significance remaining unknown.…”

A comprehensive BIAPSS platform for the physicochemical featurization of phase separating proteins

“…As a consequence (with a few exceptions [13][14][15][16] ) mostly case-by-case studies of different sequences are performed, with the broader context of many ndings, including their statistical signi cance remaining unknown.…”