Ensemble net and Protein Embeddings generalise Protein Interface Prediction beyond Homology

Author

Affiliation

Anton Feenstra

Vrije Universiteit Amsterdam, The Netherlands

Abstract

Protein interface prediction, crucial for understanding biological functions and disease mechanisms, remains a complex task in computational biology. Increasingly, Deep Learning models are having success. This study presents PIPENN-EMB which explores the added value of using embeddings from ProtT5-XL. Our results show substantial improvement over the PIPENN model for protein interface prediction, and that embeddings cover a broad range of protein features in ablation studies. PIPENN-EMB reaches state-of-the-art performance on the ZK448 benchmark dataset for protein-protein interface prediction: AUC-ROC 0.805 and MCC 0.392; where therunner-up, EnsemPPIS, reaches 0.770 and 0.291, respectively, on the same dataset. We showcase redictions on 23 resistance-related proteins from mycobacterium tuberculosis, and also compare with structure-based predictors PeSTo and CSM. Our sequence-based PIPENN-EMB method generalises to remote homologs, with less than 30% sequence identity over the whole protein sequence relative to the methods’ training dataset. In contrast, the other state-of-the-art sequence-based method, Seq-InSite and the structure-based methods PeSTo and CSM perform worse for proteins that have little recognisable homology in their training data.