PepBAN: A Deep Learning Framework with Bilinear Attention and Adversarial Learning for Peptide-Protein Interaction...

Area of Science:

• Computational biology
• Biochemistry
• Drug discovery

Background:

• Peptide-protein interactions (PepPIs) are vital for therapeutics and vaccines.
• Predicting PepPIs computationally is challenging due to limited structural and binding affinity data.
• Existing models struggle with generalization and novel peptide structures, like cyclic peptides.

Purpose of the Study:

• To develop a robust deep learning framework, PepBAN, for accurate peptide-protein interaction prediction.
• To enhance model generalization across diverse protein targets, especially with sparse binding data.
• To enable interpretation of PepPI mechanisms and address challenges in cyclic peptide interactions.

Main Methods:

• Utilized protein language model ESM-2 for protein and peptide characterization.
• Employed conditional domain adversarial learning for improved generalization.
• Developed a bilinear attention network (BAN) to model local interactions and identify key residues.
• Incorporated an atom-resolved molecular graph approach for cyclic peptides with noncanonical amino acids.

Main Results:

• PepBAN significantly outperformed state-of-the-art models on benchmark datasets.
• Demonstrated superior performance in predicting cyclic peptide-protein interactions.
• Showcased the ability to interpret interaction mechanisms through attention weight analysis.
• Successfully handled noncanonical amino acids in cyclic peptides, a key limitation of prior methods.

Conclusions:

• PepBAN offers a powerful and interpretable deep learning solution for peptide-protein interaction prediction.
• The framework advances the development of peptide-based therapeutics and vaccines.
• PepBAN provides a distinct advantage in exploring the chemical space of cyclic peptides for drug discovery.