Predicting T-cell response of presented peptides

Determine whether a peptide that is presented on the cell surface by a major histocompatibility complex (MHC) molecule will induce a T-cell response, i.e., establish accurate computational prediction of T-cell immunogenicity for presented peptides despite limited availability of T-cell assay data.

Background

The paper separates peptide vaccine design into two computational sub-tasks: predicting peptide presentation on MHC molecules and predicting whether a presented peptide induces a T-cell response. While the first task enjoys substantial data and mature predictors, the second—immunogenicity prediction—has far less experimental data from T-cell assays, making it difficult to develop reliable models.

The authors analyze heterogeneous, multi-domain T-cell response data (across peptide sources and MHC alleles) and propose a transformer-based approach with domain-aware evaluation and transfer learning techniques to mitigate shortcut learning. Despite improvements, the general challenge of predicting T-cell responses remains unresolved due to data limitations and domain heterogeneity.