Interpretable and generative deep learning models explicate phase separating intrinsically disordered motifs.

Area of Science:

• Biochemistry
• Molecular Biology
• Computational Biology

Background:

• Intrinsically disordered regions (IDRs) are crucial for protein phase separation (PS) and the organization of cellular matter into biomolecular condensates.
• Identifying specific sequence motifs and compositional features that drive PS in IDRs remains a significant challenge.

Purpose of the Study:

• To develop an interpretable deep learning framework, PhaSeMotif, for precise prediction of phase-separating motifs within IDRs.
• To experimentally validate the predicted motifs and investigate their role in PS.
• To provide a toolkit for efficient investigation of IDR motifs and insights into PS determinants.

Main Methods:

• Development of PhaSeMotif, a deep learning framework for predicting phase-separating motifs in IDRs.
• Experimental validation of predicted motifs through mutation studies to assess their impact on PS capabilities.
• Integration of generative models to create novel, validation-ready motifs.

Main Results:

• PhaSeMotif accurately predicts essential phase-separating motifs within IDRs.
• Mutations in predicted motifs significantly impair or abolish the phase separation capabilities of IDRs.
• Identified motifs exhibit diverse amino acid compositions critical for PS propensities and condensate partitioning.

Conclusions:

• PhaSeMotif offers a powerful, open-access toolkit for the efficient investigation of IDR motifs driving protein phase separation.
• The framework provides valuable insights into the molecular determinants governing PS and biomolecular condensate formation.
• The combination of prediction, generation, and validation accelerates mechanistic studies of phase-separating motifs.