A deep learning method for predicting interactions for intrinsically disordered regions of proteins.

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Intrinsically disordered proteins and regions (IDPs/IDRs) exhibit dynamic binding modes, making interface characterization difficult.
• Current structure prediction tools like AlphaFold struggle with IDR binding site prediction at standard confidence levels.

Purpose of the Study:

• To develop a novel deep-learning method, Disobind, for predicting inter-protein contact maps and interface residues involving IDRs.
• To assess Disobind's performance against state-of-the-art methods like AlphaFold-multimer and AlphaFold3.

Main Methods:

• Disobind utilizes deep learning on protein sequences to predict binding interfaces, considering the partner protein's context.
• The method does not rely on experimental structures or multiple sequence alignments.

Main Results:

• Disobind demonstrates superior performance in predicting IDR binding sites compared to AlphaFold-multimer and AlphaFold3 across various confidence thresholds.
• Combining Disobind predictions with AlphaFold-multimer further enhances prediction accuracy.

Conclusions:

• Disobind offers a robust computational approach for characterizing IDR-mediated interactions.
• The method's predictions can aid in localizing IDRs within large molecular assemblies and in modulating these interactions.