Comprehensive Evaluation and Explainable Interpretation of Peptide-HLA Binding Prediction Tools

Accurate prediction of peptide binding to human leukocyte antigen class I (HLA-I) molecules is critical for advancing immunological research, particularly in vaccine design and immunotherapy. However, limitations in model performance, interpretability, and dataset quality impede the widespread adoption of existing predictive tools. Here, we present a comprehensive evaluation of 17 HLA-I peptide binding prediction models, utilizing a meticulously curated dataset comprising over 290,000 peptides spanning 44 HLA-I alleles. We assessed model accuracy, robustness, and interpretability, employing explainability techniques such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) to elucidate underlying prediction mechanisms. Our results reveal substantial performance disparities, with self-attention-based models, including STMHCpan and BigMHC, exhibiting superior accuracy. Notably, the capsule network model CapsNet-MHC_AN demonstrated robust performance. Models trained on eluted ligand datasets outperformed those relying on binding affinity data, underscoring the critical role of high-quality training data. Ensemble and multi-algorithm approaches further improved prediction reliability. These findings highlight the need for ongoing innovation in model architecture, integration of diverse and high-quality datasets, and incorporation of structural predictors to develop more accurate, interpretable, and clinically applicable HLA-I peptide binding prediction tools.