This preprint studies a feedback loop in immunopeptidomics: predictor-dependent labels and filtering can bias the data used to train and evaluate later models. Standard discrimination metrics can remain deceptively strong even as candidate-discovery performance deteriorates on predictor-independent data.
The work reframes epitope discovery as a protein-centric learning-to-rank problem and evaluates deepMHCflare using predictor-independent data. In prospective validation, deepMHCflare ranked peptide candidates from an existing scFv immunogen; it did not design the vaccine. Two of four nominated peptides elicited significant CD8+ TNF-α+ responses. A third candidate has independent support in previous literature but was not significant in this assay.
The revised manuscript is under review. It is not an accepted journal publication.