Mapping the prediction-invisible fraction of an engineered immunopeptidome: a pilot analysis of immunoproteasome-reprogrammed mesenchymal stromal cells

Algorithms that predict which peptides the immune system sees are central to cancer vaccine design, yet they are trained almost entirely on binding affinity and remain largely blind to how the antigen-processing machinery shapes the repertoire that reaches the cell surface. Mesenchymal stromal cells engineered to express the immunoproteasome (MSC-IPr) are a whole-cell vaccine platform that presents a larger and chemically distinct peptide repertoire than dendritic cells. Here we use a pilot immunopeptidomics dataset from that platform to ask how much of a processing-reprogrammed repertoire current prediction models can see. Of 234 peptides detected only after tumour-lysate pulsing, 34 (14.5%) were empirically eluted from MHC class I yet scored as non-binders by NetMHCpan-4.1, MHCflurry affinity and MHCflurry presentation simultaneously. This is twenty-fold the rate obtained by applying identical criteria to an independently generated C57BL/6 immunopeptidome atlas (0.65%). These peptides are weak binders rather than true non-binders, and an orthogonal immunogenicity model ranks them modestly above the predictable majority. They are associated with a candidate non-canonical anchor pattern: an aspartate at P2 that is positional rather than compositional (3.4-fold above a composition-preserving null) and absent from catalogued H2-Db epitope space. Because the dataset comprises single mass-spectrometry runs without a proteasome-matched control, we present these as pilot observations, not established effects. We propose the engineered-processing strategy they illustrate, in which cells with defined, non-canonical processing serve as an empirical filter for epitope discovery, and we specify the replication, spectral-validation and control arms required to test it.