POTN-2: A deep learning-based predictor of HLA-A*02:01-restricted tumor neoantigens

Tumor neoantigens arise from somatic mutations in malignant cells and are strictly tumor-specific, minimizing autoimmune risks, making them ideal targets for T-cell immunotherapy. However, conventional pan-HLA predictors suffer from reduced accuracy and high false-positive rates for individual HLA subtypes. Our prior POTN model was trained on experimentally validated HLA-A02:01 peptide data. We developed POTN-2, an allele-specific predictor exclusively for HLA-A02:01-restricted tumor neoantigens, by expanding the feature repertoire and algorithmic architecture. We curated a comprehensive dataset comprising 50,308 experimentally confirmed positive binders, 117,334 randomly sampled negative controls, and 1012 matched mutant-wild-type neoantigen pairs. Multi-dimensional descriptors integrated, covering peptide length, residue positional encoding, amino acid composition, biological functional annotations, physicochemical properties. A three-stage feature filtering pipeline yielded three optimized subsets with 59, 65, and 109 features. We benchmarked Random Forest, Convolutional Neural Network (CNN), Support Vector Machine. The CNN trained on the full 109-feature subset achieved optimal performance, with classification accuracy of 0.936 across tumor antigens. This CNN-based POTN-2 achieved comparable performance to NetMHCpan-4.0 and DeepAntigen on tumor-associated antigen prediction, while substantially outperforming both tools in discriminating mutated neoantigen peptides from wild-type counterparts-a critical step for reducing false-positive candidates in personalized immunotherapy pipelines. We deployed an online platform (potn-2.cn)) that accepts FASTA, Excel, clinical sequencing inputs, balancing high computational throughput with subtype-specific precision to prioritize candidate neoantigens from mutation datasets. POTN-2 establishes a robust technical foundation for future extension to additional HLA alleles, and the platform adapts to emerging allele-specific models, facilitating personalized cancer immunotherapy research.