Epitope / binding prediction
Using algorithms to forecast which tumor peptides will be presented on HLA and recognized by T cells — the AI core of neoantigen discovery.
Given a tumor's mutations, the pipeline must rank potentially hundreds of candidate peptides down to the few worth putting in a vaccine. This is a stack of prediction problems: will the peptide be processed and presented? Will it bind the patient's HLA? Will a T cell recognize it (immunogenicity)?
HLA-binding prediction (e.g. NetMHCpan) is mature; immunogenicity prediction — whether presentation actually leads to a response — remains much harder and is where most current AI research, including foundation models trained on large immune datasets, is focused.
How are neoantigens predicted from a tumor biopsy?
Tumor and healthy tissue are sequenced and compared to find tumor-specific mutations; the patient's HLA type is determined; candidate mutated peptides are generated and scored for HLA binding and immunogenicity; and the top-ranked few are selected for the vaccine. AI models drive the binding and immunogenicity scoring steps.
Does a high HLA-binding score mean a peptide is immunogenic?
No — and this is a key point. Strong predicted binding only means a peptide is likely to be presented on HLA. Whether a T cell then recognizes and responds to it (immunogenicity) is a separate, much harder prediction. Many strong binders are not immunogenic, which is why binding score alone over-predicts useful targets.
Why do different epitope-prediction tools disagree with each other?
They are trained on different datasets (binding-affinity assays vs. mass-spec eluted ligands), use different model architectures, and optimize different targets. Agreement is high for strong HLA binders but drops sharply for immunogenicity, so practitioners often combine several tools rather than trusting one.