#16 · ML / AI methods

Top TCR:pMHC methods reach macro-AUC near 0.60 on unseen peptides in IMMREP25

The dominant thread is generalization of TCR:pMHC prediction to peptides absent from training data. IMMREP25 results show the strongest methods, including structural modeling approaches, reaching macro-AUC of about 0.60 on unseen peptides, a modest level. StriMap pairs structure-informed deep learning with TCR-peptide-HLA prediction and applies it to screening bacterial peptides to identify AS-linked molecular features. Alongside this, an IEDB audit finds 55.8% of MHC-I immunopeptidomics data computationally labeled, producing recursive bias; deepMHCflare reframes epitope discovery as learning-to-rank. The new emphasis is evaluation validity and label provenance alongside architecture. ViraHinter and the DefensePredictor and genomic language-model studies apply related sequence and embedding methods to virus-host and prokaryotic defense proteins, but do not address TCR specificity directly.