PHLA-SiNet: A novel peptide-HLA binding prediction model using heterogeneous Siamese neural networks

Predicting peptide-HLA binding is crucial for advancing immunotherapy; however, current models face several challenges, including peptide length variability, HLA sequence similarity, and a lack of experimentally validated negative data. To address these issues, we present PHLA-SiNet, an efficient pipeline that combines innovative representations with a lightweight architecture. PHLA-SiNet introduces three key components: (1) ESM-Pep, a peptide representation derived from a pre-trained language model (ESM), enabling flexible and training-free embedding of variable-length peptides; (2) IC-HLA, an HLA representation that captures allele-specific discriminative features using information content from binding and non-binding peptides; and (3) SiNet, a Siamese neural network that aligns peptide and HLA embeddings, bringing true binders closer in feature space. Prioritizing sensitivity-essential for reliably identifying candidate binders in early-stage immunotherapy-PHLA-SiNet is rigorously validated across benchmarks, including comparisons of IC-HLA with standard HLA encodings, analysis of SiNet's internal representations, evaluations against 11 leading predictors, and performance assessments across diverse cancer types, rare and unseen HLA types. Through computational validation using molecular docking and analysis of clinical data related to the peptide-based vaccine KIF20A-66, PHLA-SiNet demonstrates strong predictive potential in real-world contexts and highlights HLA-B08:01 as a candidate restriction element for future biological investigation. Overall, PHLA-SiNet offers a generalizable, interpretable, and resource-efficient solution for PHLA binding prediction, with strong promise to accelerate personalized immunotherapy development.