CD4+ T cell responses to complex antigens consistently focus on a handful of dominant peptide epitopes presented by major histocompatibility complex class II (MHCII) molecules, yet the rules that pick those few peptides out of thousands of possible ligands are still only partly understood. Graham et al. (Nature Medicine, 2018) used quantitative MHCII peptidomics of murine dendritic cells infected with Listeria monocytogenes to identify biochemical features of dominant epitopes and trained the deep-learning predictor BOTA (bacteria-origin T cell antigen), which was validated in mouse systems and spot-checked on a single human allele. Later work has resolved parts of the open questions that paper left: presentation-aware predictors covering diverse human HLA alleles have advanced substantially (MARIA, netMHCIIpan-4.0, MixMHC2, humanized BOTA, CAPTAn), and the underlying biology has been dissected piecemeal (influenza hemagglutinin hierarchies, cell-free processing reconstitution, HLA-DM editing studies). What remains open is the integrative core: a quantitative, mechanism-grounded account of how antigen-processing proteases, HLA-DM editing kinetics, peptide-MHCII binding stability, and antigen localization jointly select dominant epitopes, instantiated as models that predict immunodominance hierarchies of actual CD4+ T cell responses across diverse human HLA-DR, -DQ, and -DP alleles, benchmarked on allele-diverse primary-cell immunopeptidomes.

Background

CD4+ T cells recognize short peptide fragments displayed on the cell surface by major histocompatibility complex class II (MHCII; called HLA-II in humans) proteins on antigen-presenting cells such as dendritic cells. The displayed peptide repertoire is generated by the MHCII antigen-processing pathway: extracellular or endocytosed proteins are unfolded and cleaved by lysosomal proteases (chiefly cathepsins, assisted by the thiol reductase GILT), and the resulting fragments compete for binding into the open groove of MHCII molecules. The nonclassical HLA-DM molecule acts as a peptide-exchange catalyst (an 'editor') that accelerates the release of weakly bound peptides, so the final repertoire is biased toward high-stability complexes; the residence time, or off-rate, of a peptide bound to MHCII is a key kinetic parameter of this selection. A central, long-standing observation is immunodominance: even though a pathogen or commensal genome encodes thousands of peptides that can bind MHCII, the CD4+ T cell response concentrates on a few dominant epitopes whose relative response magnitudes form an immunodominance hierarchy. Deciding which epitopes will dominate matters directly for vaccine target selection, cancer neoantigen design, and understanding autoimmunity, and the determinants have been reviewed for decades (e.g., Kim and Sadegh-Nasseri, 2015): abundance and localization of the source protein, protease cleavage preferences, MHCII binding affinity and complex stability, HLA-DM editing, and the available T cell receptor (TCR) repertoire all contribute, but their relative weights have resisted quantification.

Unbiased measurement became possible with immunopeptidomics: immunoaffinity purification of MHCII molecules followed by mass spectrometry identifies the natural ligand repertoire of a cell without guessing candidate antigens. Graham et al. (2018) applied a quantitative version of this method to bone-marrow-derived murine dendritic cells (carrying the MHCII allele I−AbI-A^b) exposed to live Listeria monocytogenes, identified four dominant bacterial CD4+ T cell epitopes and established their hierarchy in vivo, and profiled over 3,600 endogenous mouse peptides. From these data they distilled composite features of dominant epitopes—derivation from secreted proteins or extracellular regions of cell-wall proteins, conformity to the MHCII binding motif, distance of more than eight amino acids from transmembrane domains, and tertiary-structure accessibility to proteases and the MHCII groove—and trained BOTA (bacteria-origin T cell antigen), a deep neural network that predicts immunodominant epitopes from an annotated genome alone. Functional validation combined single-cell TCR sequencing with DNA-encoded peptide-MHCII library screening to confirm recognition by responding T cells. Crucially for the present question, the model was trained on murine peptidomics; the only human assessment was recovery of IEDB (Immune Epitope Database) curated epitopes for the inflammatory-bowel-disease risk allele HLA-DRB1*01:03 (38 of 50, with 12 known epitopes missed). The paper closes by framing the obstacles that remain: the mechanistic determinants of immunodominance are 'incompletely understood', and genome-scale, presentation-aware prediction had not been extended or validated across diverse human HLA alleles.

Subsequent work built a strong presentation-prediction toolkit. MARIA (Chen et al., 2019) is a recurrent neural network trained on mass-spectrometry ligand data, gene-expression levels, and protease-cleavage signatures; netMHCIIpan-4.0 (Reynisson et al., 2020) added eluted-ligand predictors across many HLA alleles by motif deconvolution of immunopeptidomes; MixMHC2 (Racle et al., 2023) extended ligand prediction to essentially any MHC-II allele and exposed an unexpected reverse binding mode in HLA-DP; and the same group that created BOTA humanized it (hBOTA; Pedersen et al., 2022) and then built CAPTAn (Stražar et al., 2023) on 358,024 monoallelic HLA-II ligands (ligands eluted from cell lines each expressing a single HLA-II allele) with deliberate coverage of the understudied HLA-DQ and HLA-DP loci. In parallel, mechanistic studies advanced on individual systems: the human CD4 immunodominance hierarchy of influenza hemagglutinin was shown to arise by selection of the memory repertoire toward peptides likely to be processed and presented by dendritic cells (Cassotta et al., 2020); a cell-free processing system reconstituting HLA-DR, HLA-DM, and cathepsins reproduced epitope selection for HIV-1 proteins and showed it depends on protein stability and solvent accessibility (Sengupta et al., 2023); and a reconstituted-processing-plus-prediction framework across 11 HLA-DRB1 allotypes showed that SARS-CoV-2 dominant epitopes arise by antigen-dependent processing routes (Álvaro-Benito et al., 2025). HLA-DM/DO editing remains an active mechanistic focus (Welsh and Sadegh-Nasseri, 2020). Yet these threads have not been unified: current predictors are scored by how well they recover presented ligands, not by how accurately they forecast the ranking and magnitude of T cell responses (immunodominance), and no published model quantitatively combines processing kinetics, DM editing, complex stability, and antigen localization to predict hierarchies across human HLA diversity.

Problem Statement

The problem, as left open by Graham et al. 2018 and reframed by the intervening literature, is the quantitative specification of immunodominance hierarchies for MHCII-restricted CD4+ T cell responses: given an antigen (or a genome) and a human HLA class II context, predict which few epitopes will dominate the CD4+ T cell response and their approximate ranking, and explain quantitatively why. It has two jointly supporting components. (1) Mechanistic: determine how the individual determinants—processing by lysosomal proteases and reductases, HLA-DM-mediated peptide editing kinetics, peptide-MHCII binding stability (off-rates), and the localization of the source protein in secreted or accessible extracellular regions—interact and with what quantitative weights to select the narrow dominant set out of the many peptides that can bind MHCII. (2) Predictive and validating: build presentation-aware, accessibility-aware models of the BOTA type that work genome-scale across diverse human HLA-DR, HLA-DQ, and HLA-DP alleles, benchmarked on allele-diverse, primary-cell-derived human immunopeptidomes and against measured T cell responses, closing the recovery gap for known epitopes that current tools miss. Component (1) supplies the mechanism that makes component (2)'s predictions explanatory and generalizable rather than correlative, and component (2) supplies the genome-scale, cross-allele test that decides whether component (1)'s quantitative account is complete. The host-commensal biological consequences of recognizing such epitopes (e.g., T cells specific for SusC, a TonB-dependent outer-membrane receptor of commensal Bacteroides species, that produce the regulatory cytokine IL-10 in homeostasis versus disease) and the engineering throughput of TCR-peptide-MHCII validation are explicitly excluded as separate research programs also named by the source, not parts of this problem's completion boundary. The available T cell receptor (TCR) repertoire is held fixed rather than modelled: the four presentation-side determinants named in component (1) are the intended closed set, and an answer that predicts the immunodominance hierarchy from those determinants without modelling repertoire selection lies within the completion boundary.

References

  1. Graham DB, Luo C, O'Connell DJ, Lefkovith A, Brown EM, Yassour M, Varma M, Abelin JG, Conway KL, Jasso GJ, Matar CG, Carr SA, Xavier RJ. Antigen discovery and specification of immunodominance hierarchies for MHCII-restricted epitopes. Nature Medicine 24(11):1762-1772 (2018). DOI: 10.1038/s41591-018-0203-7. https://doi.org/10.1038/s41591-018-0203-7
  2. Pedersen TK, Brown EM, Plichta DR, Johansen J, Twardus SW, Delorey TM, Lau H, Vlamakis H, Moon JJ, Xavier RJ, Graham DB. The CD4+ T cell response to a commensal-derived epitope transitions from a tolerant to an inflammatory state in Crohn's disease. Immunity 55(10):1909-1923.e6 (2022). DOI: 10.1016/j.immuni.2022.08.016. https://doi.org/10.1016/j.immuni.2022.08.016
  3. Stražar M, Park J, Abelin JG, Taylor HB, Pedersen TK, Plichta DR, Brown EM, Eraslan B, Hung YM, Ortiz K, Clauser KR, Carr SA, Xavier RJ, Graham DB. HLA-II immunopeptidome profiling and deep learning reveal features of antigenicity to inform antigen discovery. Immunity 56(7):1681-1698.e13 (2023). DOI: 10.1016/j.immuni.2023.05.009. https://doi.org/10.1016/j.immuni.2023.05.009
  4. Chen B, Khodadoust MS, Olsson N, Wagar LE, Fast E, Liu CL, Muftuoglu Y, Sworder BJ, Diehn M, Levy R, Davis MM, Elias JE, Altman RB, Alizadeh AA. Predicting HLA class II antigen presentation through integrated deep learning. Nature Biotechnology 37(11):1332-1343 (2019). DOI: 10.1038/s41587-019-0280-2. https://doi.org/10.1038/s41587-019-0280-2
  5. Reynisson B, Alvarez B, Paul S, Peters B, Nielsen M. NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS-derived eluted ligand data. Nucleic Acids Research 48(W1):W449-W454 (2020). DOI: 10.1093/nar/gkaa379. https://doi.org/10.1093/nar/gkaa379
  6. Racle J, Guillaume P, Schmidt J, Michaux J, Larabi A, Lau K, Perez MAS, Croce G, Genolet R, Coukos G, Zoete V, Pojer F, Bassani-Sternberg M, Harari A, Gfeller D. Machine learning predictions of MHC-II specificities reveal alternative binding mode of class II epitopes. Immunity 56(6):1359-1375.e13 (2023). DOI: 10.1016/j.immuni.2023.03.009. https://doi.org/10.1016/j.immuni.2023.03.009
  7. Cassotta A, Paparoditis P, Geiger R, Mettu RR, Landry SJ, Donati A, Benevento M, Foglierini M, Lewis DJM, Lanzavecchia A, Sallusto F. Deciphering and predicting CD4+ T cell immunodominance of influenza virus hemagglutinin. Journal of Experimental Medicine 217(10):e20200206 (2020). DOI: 10.1084/jem.20200206. https://doi.org/10.1084/jem.20200206
  8. Sengupta S, Zhang J, Reed MC, Yu J, Kim A, Boronina TN, Board NL, Wrabl JO, Shenderov K, Welsh RA, Yang W, Timmons AE, Hoh R, Cole RN, Deeks SG, Siliciano JD, Siliciano RF, Sadegh-Nasseri S. A cell-free antigen processing system informs HIV-1 epitope selection and vaccine design. Journal of Experimental Medicine 220(7):e20221654 (2023). DOI: 10.1084/jem.20221654. https://doi.org/10.1084/jem.20221654
  9. Álvaro-Benito M, Abualrous ET, Lingel H, Meltendorf S, Holzapfel J, de Diego Valera P, Sticht J, Kuropka B, Clementi C, Kuppler F, Brunner-Weinzierl MC, Freund C. Cut or bind? Antigen-specific processing mechanisms define CD4+ T cell immunodominant epitopes for SARS-CoV-2 S and N proteins. Genome Medicine 17(1):147 (2025). DOI: 10.1186/s13073-025-01577-8. https://doi.org/10.1186/s13073-025-01577-8
  10. Welsh RA, Sadegh-Nasseri S. The love and hate relationship of HLA-DM/DO in the selection of immunodominant epitopes. Current Opinion in Immunology 64:117-123 (2020). DOI: 10.1016/j.coi.2020.05.007. https://doi.org/10.1016/j.coi.2020.05.007
  11. Kim A, Sadegh-Nasseri S. Determinants of immunodominance for CD4 T cells. Current Opinion in Immunology 34:9-15 (2015). DOI: 10.1016/j.coi.2014.12.005. https://doi.org/10.1016/j.coi.2014.12.005
  12. Kohlgruber AC, Dezfulian MH, Sie BM, Wang CI, Kula T, Laserson U, Larman HB, Elledge SJ. High-throughput discovery of MHC class I- and II-restricted T cell epitopes using synthetic cellular circuits. Nature Biotechnology 43(4):623-634 (2024). DOI: 10.1038/s41587-024-02248-6. https://doi.org/10.1038/s41587-024-02248-6