Drastic improvement in MHC class I epitope prediction through training data augmentation, generative adversarial network modeling of selfness, and consideration of homology to class II epitopes.
A bioRxiv preprint introduces DeepOmicsNeo, a modular deep learning framework combining three components: a binding module that synthesizes pseudo-binders and non-binders by editing non-anchor residues to overcome data scarcity in non-9mer peptide–MHC class I prediction; a 'selfness' module using a GAN trained on approximately 60 million human self-peptides to quantify similarity to self in an unsupervised manner, with scores correlating with IFN-gamma secretion by T cells; and a module exploiting central-region homology between HLA class I and class II epitopes. The integrated immunogenicity model reportedly outperforms state-of-the-art predictors in benchmark testing. Why it matters: high false-positive rates in current predictors are a key rate limiter for personalized cancer vaccines, and a better-ranking immunogenicity model could raise the hit rate of selected targets without adding clinical cost. It is method-stage and unreviewed — no prospective or clinical validation is reported.