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

Area of Science:

• Biochemistry
• Molecular Biology
• Computational Biology

Background:

• Liquid-liquid phase separation (LLPS) is a crucial mechanism for forming membraneless organelles, regulating cellular physiology, and is implicated in amyloid formation.
• Proteins involved in LLPS are categorized as scaffolds and clients, yet computational prediction of client proteins from amino acid sequences is underdeveloped.

Purpose of the Study:

• To develop an accurate computational method for predicting liquid-liquid phase separation (LLPS) client proteins from amino acid sequences.
• To identify novel LLPS client proteins and analyze sequence properties differentiating scaffolds from clients.

Main Methods:

• Utilized a deep-learning technique, Transformer, to extract information-rich features from amino acid sequences.
• Employed supervised machine learning for the prediction of LLPS client proteins.
• Validated predictions by assessing known LLPS regulators and localization enrichment in membraneless organelles.

Main Results:

• Seq2Phase demonstrated high accuracy in predicting LLPS client proteins across multiple species (human, mouse, yeast, plant).
• Feature analysis revealed distinct sequence properties between LLPS scaffolds and clients, highlighting limitations in current knowledge.
• The study predicts a substantial number of undiscovered LLPS client proteins, suggesting broad applicability of the method.

Conclusions:

• Seq2Phase is an accurate and broadly applicable tool for identifying LLPS client proteins.
• The findings suggest numerous undiscovered LLPS client proteins exist, necessitating further research.
• This work will enhance understanding of LLPS's molecular basis, physiological roles, and disease implications.