The study team began with a panel of peptides enriched for likely MHC ligands. These peptides were derived from substrates of Mycobacterium tuberculosis’ type VII secretion systems (T7SS) and other candidate vaccine proteins that emerged as HLA-A*02:01 binders in a yeast display screen. Investigators then infected human macrophages with Mtb and used targeted mass spectrometry (MS) to search for those specific peptides on MHC-I molecules eluted from infected cells. The result was strikingly sparse: fewer than 1% of the yeast-display–identified binders were detected at measurable levels on MHC-I from infected macrophages. This mismatch indicates that biochemical MHC binding, as measured in yeast, does not by itself predict presentation during infection.
The authors reported these detection results without providing additional experimental parameters in the abstract; the source did not list peptide numbers, MS sensitivity thresholds, or detailed infection conditions in that section. What the data clearly indicate is a pronounced gap between in vitro binding potential and in situ antigen presentation.
To probe why most binders were not observed on infected cells, the investigators applied a computational modeling strategy based on neural networks. The model incorporated established MHC binding information plus other biological features. Two additional predictors emerged as informative: transcript or gene expression level and the identity of amino acids adjacent to the candidate epitope, which may influence proteolytic processing. In the authors’ words, these features distinguished peptides that were both MHC binders and actually presented on infected cells from those that remained undetected.
The preprint abstract does not enumerate the precise neural network architecture, training set size, or performance metrics here, and the source did not present those implementation details in the abstract text. Still, the approach clarifies that antigen presentation in the context of infection reflects multiple molecular constraints beyond straightforward peptide-HLA affinity.
After integrating expression and flanking-sequence features into their prediction pipeline, the team returned to targeted MS as the experimental readout. Applying these refined criteria enabled detection of additional novel MHC-I epitopes from Mtb-infected macrophages that had not been captured by simple HLA-binding screens alone. This iterative coupling of computational prioritization and sensitive MS readout demonstrates a practical workflow: start with broad binder screens, add infection-relevant filters, then validate by targeted proteomics.
The abstract does not supply the exact peptide sequences identified in the follow-up MS nor the total count of new epitopes discovered; those specifics may be in the full text or supplementary materials. The key message is methodological: adding biologically relevant predictors improves the yield of infection-relevant epitope discovery.
Data indicated that peptides presented on MHC-I by Mtb-infected cells constitute a very small fraction of all theoretically possible MHC-binding peptides derivable from the pathogen proteome. The authors interpreted this as the outcome of multiple sequential bottlenecks: antigen expression within the infected cell, proteolytic generation of the correct peptide termini (influenced by adjacent amino acids), peptide transport and loading onto MHC, and the stability of peptide–MHC complexes in the infected-cell context. In short, MHC affinity is necessary but not sufficient for presentation during infection.
This nuance matters for immune monitoring and vaccine design. Approaches that rely solely on biochemical binding assays are likely to nominate many candidates that will not be presented during natural infection. The investigators therefore suggest incorporating expression and processing context into antigen-selection pipelines when the goal is to prime CD8+ T cell responses against intracellular bacterial pathogens such as Mycobacterium tuberculosis.
The article is a preprint and has not been certified by peer review. The authors declared that O.L., B.D.B., and F.M.W. are inventors on a patent application related to tuberculosis vaccine designs; funders are listed in the source. The abstract does not report detailed experimental parameters—such as numbers of peptides tested, macrophage donor variability, quantitative MS detection limits, or the neural network’s performance metrics—so readers should consult the full manuscript and supplementary files for those specifics. Finally, the authors frame their conclusions as informing antigen-prediction models and vaccine antigen selection strategies rather than as definitive clinical recommendations.
Open questions documented or implied in the source that remain to be addressed include how generalizable the identified predictive features are across HLA alleles, different host cell types, and diverse Mtb strains, and how assay sensitivity and infection conditions affect the observed fraction of presented peptides. These are the concrete next steps and limitations reported in the preprint.