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

Intrinsically disordered proteins or regions (IDPs or IDRs) exist as ensembles of conformations in the monomeric state and can adopt diverse binding modes, making their experimental and computational characterization challenging. Here, we developed Disobind, a deep-learning method that predicts inter-protein contact maps and interface residues for an IDR and a partner protein, leveraging sequence embeddings from a protein language model. Several current methods, in contrast, provide partner-independent predictions, require the structure of either protein, and/or are limited by the MSA quality. Disobind performs better than AlphaFold-multimer and AlphaFold3. Combining the Disobind and AlphaFold-multimer predictions further improves the performance. However, Disobind is limited to binary IDP-partner complexes, where the two proteins are known to bind, and the input sequence fragments are less than one hundred residues long. The predictions can be used to localize IDRs in integrative structures of large assemblies, characterize protein-protein interactions involving IDRs, and modulate IDR-mediated interactions.