Seq2Phase: language model-based accurate prediction of client proteins in liquid–liquid phase separation

Abstract

Abstract Motivation Liquid–liquid phase separation (LLPS) enables compartmentalization in cells without biological membranes. LLPS plays essential roles in membraneless organelles such as nucleoli and p-bodies, helps regulate cellular physiology, and is linked to amyloid formation. Two types of proteins, scaffolds and clients, are involved in LLPS. However, computational methods for predicting LLPS client proteins from amino-acid sequences remain underdeveloped. Results Here, we present Seq2Phase, an accurate predictor of LLPS client proteins. Information-rich features are extracted from amino-acid sequences by a deep-learning technique, Transformer, and fed into supervised machine learning. Predicted client proteins contained known LLPS regulators and showed localization enrichment into membraneless organelles, confirming the validity of the prediction. Feature analysis revealed that scaffolds and clients have different sequence properties and that textbook knowledge of LLPS-related proteins is biased and incomplete. Seq2Phase achieved high accuracies across human, mouse, yeast, and plant, showing that the method is not overfitted to specific species and has broad applicability. We predict that more than hundreds or thousands of LLPS client proteins remain undiscovered in each species and that Seq2Phase will advance our understanding of still enigmatic molecular and physiological bases of LLPS as well as its roles in disease. Availability and implementation The software codes in Python underlying this article are available at https://github.com/IwasakiLab/Seq2Phase.