GeoPMB: An Interface-Aware Geometric Deep Learning Framework for Peptide-MHCI Binding Prediction with Evolutionary...

The binding of peptides to class I major histocompatibility complex (MHCI) molecules is central to adaptive immunity, making the identification of immunogenic peptides a critical step in developing effective vaccines and immunotherapies against bacterial, viral, and even cancer-related diseases. However, given that experimental determination of immunogenicity is extremely resource-intensive and time-consuming, developing accurate computational methods is indispensable for high-throughput prediction. Existing sequence-based predictors are effective in data-rich regimes but often struggle to generalize to unseen or less-explored alleles and lack biophysical interpretability. Moreover, current structure-based methods frequently fail to capture fine-grained 3D geometric constraints essential for precise binding prediction. To address these limitations, we propose GeoPMB, a novel framework that synergistically integrates geometric deep learning with pretrained protein language models (PLMs). GeoPMB leverages PLMs to extract evolutionary contexts and functional semantics from sequences, while employing a geometric graph network to explicitly model the spatial dependencies and interfacial physicochemical features within the predicted pMHCI complex. Extensive benchmarks demonstrate that GeoPMB consistently outperforms state-of-the-art baselines, exhibiting superior generalization in pMHCI binding specificity and affinity prediction, particularly for rare alleles. Furthermore, GeoPMB shows remarkable versatility, achieving high performance in antibody-antigen docking pose ranking. These results highlight the potential of GeoPMB as a powerful, structurally aware tool for precision immunology and therapeutic design.