Improving generalizability for MHC-I binding peptide predictions through structure-based geometric deep learning

Abstract

AbstractUnderstanding MHC peptide presentation is crucial for pathogen recognition, autoimmune disease treatment, and cancer immunotherapy development.In silicoprediction of MHC-bound peptides is essential for cost-effective therapy design. However, state-of-the-art sequence-based (SeqB) methods encounter challenges in sensitivity to data biases and limited generalizability, particularly for less-studied MHC alleles. We hypothesize that structure-based (StrB) methods can enhance generalization by leveraging encoded physics and geometric rules. Introducing three supervised StrB geometric deep learning (GDL) approaches, we demonstrate their superior generalization outperforming two SeqB methods by 5 to 11% in AUC. To enhance data efficiency, we present a self-supervised learning approach, 3D-SSL, surpassing SeqB methods without using binding affinity data. We demonstrated StrB method resilience to biases in binding data using a case study on HBV vaccine design. These findings emphasize the capacity of structure-based methods to enhance generalizability and efficiently use limited data, bearing implications for data-intensive fields like T-cell receptor specificity predictions.