Immunoproteasome-Reprogrammed MSCs Reveal a Blind Spot in Epitope Prediction
Binding-based epitope prediction captures only part of what antigen-processing cells actually present. In this pilot immunopeptidomics analysis, immunoproteasome-reprogrammed mesenchymal stromal cells generated a lysate-induced peptide set that current algorithms largely failed to recognize, even when those peptides were experimentally eluted from MHC class I. The finding matters because the field still leans heavily on sequence affinity models to choose vaccine targets, yet those models are trained mainly on binding chemistry rather than on the upstream proteolytic decisions that determine which fragments ever reach the cell surface.
The dataset was small, but the signal was not subtle. Of 234 peptides detected only after tumour-lysate pulsing, 34 peptides, or 14.5%, were experimentally eluted from MHC class I while scoring as non-binders by NetMHCpan-4.1, MHCflurry affinity, and MHCflurry presentation all at once. When the same criteria were applied to an independently generated C57BL/6 immunopeptidome atlas, the rate fell to 0.65%. The gap was wide enough to suggest that the engineered antigen-processing state of these cells exposes a category of ligands that conventional prediction pipelines systematically miss.
One sentence captures the core practical problem: the immune system can see peptides that the software does not.
A Distinct Processing Program Produces a Distinct Peptide Repertoire
The experimental platform is central to interpreting the result. Mesenchymal stromal cells were engineered to express the inducible immunoproteasome subunits β1i, β2i, and β5i, creating a mixed proteasome population rather than a pure immunoproteasome pool. Those cells were pulsed with EL4 tumour lysate and compared with matched unpulsed controls, while dendritic cells underwent the same treatment. The study then partitioned the detected ligands into lysate-induced and constitutively presented compartments on the basis of differential detection.
This is not a trivial bookkeeping exercise. A peptide appearing only in the pulsed condition may reflect altered proteolysis, altered loading, altered detection sensitivity, or a combination of all three. The analysis therefore stays appropriately cautious and treats the observations as pilot-stage associations. Even with that restraint, the lysate-induced compartment of the immunoproteasome-reprogrammed cells appears to contain a qualitatively unusual set of ligands, including peptides with weak but real binding signatures that fell below the threshold of standard prediction models.
The functional implication is straightforward: cells with a defined non-canonical processing program can act as an empirical filter for epitope discovery. In other words, rather than asking only which peptides ought to bind, the system asks which peptides are actually generated by a specialized proteolytic environment and then displayed.
Weak Binders, Not True Non-Binders, Fill the Missing Space
The peptides missed by all three prediction tools were not random debris. They were described as weak binders rather than true non-binders, and an orthogonal immunogenicity model ranked them modestly above the predictable majority. That distinction is biologically plausible. Binding predictors are calibrated to canonical HLA motif space, so peptides generated by unusual cleavage patterns or by non-canonical trimming may look invisible at the first pass even when they still possess enough binding energy to sit in the MHC groove.
This is exactly the kind of failure mode expected when antigen processing is reprogrammed upstream of binding. The immunoproteasome changes cleavage preference, peptide length distribution, and the positional exposure of anchor residues. If a peptide is produced by a route that does not match the statistical pattern embedded in training data, affinity-based ranking can undercall it even when the ligand is real. In a vaccine context, that means a tumor fragment can be discarded as unpromising before the relevant immunobiology has even been tested.
A single peptide class illustrates the point particularly well.
A Candidate Non-Canonical Anchor Pattern Emerged at P2
Among the prediction-invisible peptides, the analysis identified a candidate non-canonical anchor pattern: an aspartate at P2 that was positional rather than compositional, appearing 3.4-fold above a composition-preserving null and absent from catalogued H2-Db epitope space. In plain terms, the residue was not just present somewhere in the sequence; it was enriched in a specific position that matters for MHC binding geometry. If confirmed, this would represent a binding motif not well represented in the current reference epitope universe.
The biological relevance is easy to see. Anchors at defined positions determine how a peptide sits in the groove, and a shifted preference can reshape the entire immunopeptidome. A non-canonical anchor pattern also helps explain why standard predictors, which are trained on prevailing motif conventions, may struggle to score these ligands correctly. The result does not prove that every peptide with this feature is immunogenic, but it does argue that the current map of H2-Db-presented epitopes is incomplete.
Why the Pilot Design Still Needs Hard Validation
Several design features keep this interpretation firmly within pilot territory. The analysis relied on single mass-spectrometry runs, and there was no proteasome-matched control. The comparison between MSC-IPr and dendritic cells is therefore vulnerable to cell-type and batch confounding, even though the within-MSC-IPr comparison is more insulated because both lysate-induced and constitutively presented compartments were drawn from the same pulsed run. The authors also make clear that empty-vector MSCs were not profiled by mass spectrometry in the source experiment, limiting direct attribution of the observed repertoire shift to the engineered immunoproteasome alone.
Still, the study is valuable because it defines the next experimental step with unusual clarity. Replication in independent runs, spectral validation, and control arms matched for proteasome state would be required to separate genuine processing effects from sampling noise. A proteasome-matched control would be especially important for determining whether the invisible fraction is truly driven by the immunoproteasome reprogramming or by another aspect of the cellular response to lysate pulsing.
The bedside relevance sits downstream of those technical details. Cancer vaccine design depends on finding ligands the tumor actually presents, not only the ligands that fit a consensus binding model. When an engineered antigen-processing platform surfaces peptides that are experimentally real but computationally invisible, it exposes a gap in current neoepitope discovery pipelines and argues for broader use of empiric, cell-based discovery systems alongside in silico ranking.
For now, the main conclusion is narrower than any headline claim and more useful to the field: the prediction machinery misses a measurable fraction of ligands produced by an engineered immunopeptidome, and the miss is enriched for peptides with unusual anchor geometry. Whether that fraction persists under independent replication, and whether it translates into better vaccine targets, remains a question for the next round of experiments.