pHLA-Bi-Mamba: A pan-specific deep learning model for peptide-HLA-Ⅰ binding affinity and eluted ligand prediction

Human leukocyte antigen class I (HLA-I) molecules are central to immunosurveillance, and accurate prediction of peptide-HLA-I binding is critical for immune target discovery and personalized immunotherapies. Although Transformer-based deep learning models have advanced sequence representation learning, they are often limited by quadratic computational complexity. In contrast, the bidirectional Mamba architecture integrates forward and backward contextual information, enabling more effective modeling of global sequence dependencies and spatial constraints within the peptide-binding groove. Because peptide-MHC binding is governed by global physicochemical interactions-where N- and C- terminal residues are frequently thermodynamically coupled−such bidirectional context is particularly advantageous. Here, we present pHLA-Bi-Mamba, a pan-specific model that leverages bidirectional Mamba-based protein language modeling to predict both binding affinity and eluted ligands for peptide-HLA-I pairs. pHLA-Bi-Mamba achieves state-of-the-art performance on recent benchmarking datasets while providing interpretable attribution to key peptide and HLA positions. For binding affinity prediction, the model attains an R 2 of 0.220, substantially outperforming NetMHCpan-4.1 (0.107) and MHCflurry-2.0 (−0.004). For eluted ligand prediction, it improves the area under the precision–recall curve (AUPRC) by at least 20% on an independent IEDB test set. To our knowledge, this work represents the first application of the bidirectional Mamba architecture to peptide-HLA-I binding prediction. pHLA-Bi-Mamba provides an efficient tool for large-scale immune epitope screening, particularly for identifying rare binders within highly imbalanced datasets. The source code and curated datasets are available at GitHub: https://github.com/ImmunoInformatics-dev/pHLA-Bi-Mamba .