Immunological validation of AI-assisted neoantigen prioritization: from peptide-HLA presentation to functional T-cell evidence

Artificial intelligence and machine learning models increasingly determine which tumour mutations proceed to experimental testing and individualized vaccine manufacture. The labels used to train and evaluate these models, however, represent different immunological events: peptide-HLA binding, antigen processing and presentation, T-cell-receptor engagement, T-cell effector function, vaccine-induced immunity and clinical benefit. Treating these events as interchangeable obscures where immune evidence ends and untested inference begins. We curated an evidence map of 32 predictive or generative models and classified the immunological evidence in each model’s first report. All 32 reported computational performance. Eight generated new experimental peptide-HLA data, and four tested candidates that the model itself had selected or designed. Three of the four applied a T-cell functional readout to their own predictions. One of the three did so for tumour neoantigens. None evaluated model output as a prespecified tested variable in a study of vaccine-induced immunity or clinical outcome. Excluding three genomic HLA-disruption models left the pattern unchanged among 29 peptide-level models. We interpret the gap as an immunological measurement problem. Peptide-HLA presentation, T-cell recognition and effector function require different assays, controls and negative definitions, while global discrimination metrics do not reproduce patient-level selection under a limited validation budget. We map model claims to the immune events and readouts that can test them. We separate successful vaccine pipelines from prospective validation of the algorithms that selected their antigens. We then propose assay-matched, model-locked studies with patient-level functional endpoints. The counts are marginal, not nested, and they describe this curated, non-systematic set rather than model quality or the prevalence of validation practice across the field. The first-report rule undercounts mature models validated in later studies. AI-assisted neoantigen prioritization will become immunologically informative when each computational claim is tested at the biological transition it is intended to predict.