Unveiling immune targets via TriStageHLA: high-throughput screening of peptide presentation, peptide–HLA binding, and immunogenicity

High-throughput screening of HLA class I epitopes underpins neoantigen vaccine design, T cell therapies, and pathogen immune monitoring, yet proteome-scale prioritization remains challenging under severe class imbalance. We present TriStageHLA, a compact three-stage framework that decomposes epitope discovery into peptide-only screening (PRE), allele-specific binding/presentation prediction (BIND), and immunogenicity prediction (IM). PRE performs early pruning via a LoRA-adapted ESM-2 classifier; BIND integrates quantitative binding affinity and eluted ligand data with a dual-branch architecture combining global cross-attention and local motif extraction; IM is fine-tuned on curated human T-cell assay data with a decoupled LoRA head. Trained on IEDB-derived datasets with per-allele balancing and allele-discordant supervision, TriStageHLA-BIND achieves AUROC of 96.2% and AUPRC of 95.2% on combined BA/EL evaluation, while TriStageHLA-IM reaches AUROC of 81.4\% and AUPRC of 73.8\% on held-out immunogenicity data, improving over the strongest immunogenicity baselines ImmuneApp (70.7% / 63.6%) and BigMHC-IM (60.1% / 51.5%). With substantially fewer trainable parameters than prior models, TriStageHLA enables efficient, calibrated epitope prioritization across binding, presentation, and immunogenicity. Code and pretrained weights are available at https://github.com/SkywalkerLuke/TriStageHLA.