An interpretable peptide-HLA model emergently learns binding energetics and structure
- 1Duke University School of Medicine
Abstract
The range of peptides a human leukocyte antigen (HLA) binds and displays modulates immune response and therefore underpins vaccine design, neoantigen discovery, autoimmunity, transplantation, and hypersensitivity reactions. Modern predictors of peptide-HLA (pMHC) binding and presentation are remarkably accurate, but they are black boxes; their internal computations are opaque and post-hoc explanatory methods lack guarantees of attribution. Here we introduce LAtent Motif INteraction Aggregation (LAMINA), an architecture whose prediction is, by construction, interpretable and attributable. LAMINA embeds every gapless sub-sequence of the HLA pseudosequence and of the candidate peptide as a learned "soft" motif, scores every HLA-peptide-motif pair, and lastly aggregates those scores to produce a prediction. Despite having only 4.7 million parameters and training in about ten hours on a single desktop workstation, LAMINA matches or exceeds state-of-the-art predictors on held-out binding-affinity regression and is competitive on rank correlation. Strikingly, the model states correlate with interaction energies and other structural metrics in structurally characterized pMHC complexes, despite training exclusively on sequence data alone.
Related
- XAMPGen: An Interpretable SI Pipeline For Novel Antimicrobial Peptide Discovery
- Accelerating Molecular Dynamics with Hydrogen Mass Repartitioning to Prioritize Immunogenic Neoantigens for Cancer Vaccines through Peptide HLA Binding Free Energy Calculations
- Abdominal-B regulates the male seminal fluid transferome and new fecundity factors required for sperm sex peptide binding
- ToxINsight: An Integrated Web Platform for Sequence, Structure, and Ligand Binding Analysis of Snake Venom Toxins
- scFLAME: a unified generative model for interpretable clustering, hierarchical structure discovery and marker-gene identification in single-cell RNA-seq data
BibTeX
@misc{chen2026interpretable,
title = {{An interpretable peptide-HLA model emergently learns binding energetics and structure}},
author = {S. F. Chen and R. J. Steele and E. K. Oermann},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.08.750216},
url = {https://www.biorxiv.org/content/10.64898/2026.09.08.750216v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - An interpretable peptide-HLA model emergently learns binding energetics and structure
AU - S. F. Chen
AU - R. J. Steele
AU - E. K. Oermann
PY - 2026
DA - 2026/09/08
PB - bioRxiv
DO - 10.64898/2026.09.08.750216
UR - https://www.biorxiv.org/content/10.64898/2026.09.08.750216v1
AB - The range of peptides a human leukocyte antigen (HLA) binds and displays modulates immune response and therefore underpins vaccine design, neoantigen discovery, autoimmunity, transplantation, and hypersensitivity reactions. Modern predictors of peptide-HLA (pMHC) binding and presentation are remarkably accurate, but they are black boxes; their internal computations are opaque and post-hoc explanatory methods lack guarantees of attribution. Here we introduce LAtent Motif INteraction Aggregation (LAMINA), an architecture whose prediction is, by construction, interpretable and attributable. LAMINA embeds every gapless sub-sequence of the HLA pseudosequence and of the candidate peptide as a learned "soft" motif, scores every HLA-peptide-motif pair, and lastly aggregates those scores to produce a prediction. Despite having only 4.7 million parameters and training in about ten hours on a single desktop workstation, LAMINA matches or exceeds state-of-the-art predictors on held-out binding-affinity regression and is competitive on rank correlation. Strikingly, the model states correlate with interaction energies and other structural metrics in structurally characterized pMHC complexes, despite training exclusively on sequence data alone.
ER -