Learning with Privileged Knowledge Distillation for Improved Peptide–Protein Docking

Protein–peptide interactions play a key role in many biological processes and the development of new therapeutics. To develop a deeper understanding of the intricacies driving these interactions, it is crucial to structurally characterize them at the atomic level. As the experimental determination of such complexes is often difficult and costly, various computational methods for modeling the structures of receptor–peptide complexes have been developed. Among these, AlphaFold-Multimer (AFM) is one of the top-performing complex structure prediction methods. However, despite their generally high accuracy in modeling protein complexes, models of protein–peptide complexes often have significant errors. Here, we present DistPepFold, which improves protein–peptide complex docking using an AFM-based architecture through a privileged knowledge distillation approach. DistPepFold leverages a teacher model trained with native interaction information, which then transfers its knowledge to a student model via a teacher–student distillation process. We evaluated DistPepFold on two data sets and showed that DistPepFold outperforms AFM and other existing peptide docking methods.