MHChron: diversity-balanced dataset design for robust peptide–MHC binding prediction across MHC class I and II

Description This Zenodo archive accompanies the GitHub repository for MHChron, a peptide–MHC binding prediction framework, and provides the large data files omitted from the repository due to their size. The corresponding GitHub repository contains the complete source code used for data collection, processing, clustering, feature generation, model training, evaluation, benchmarking against external methods, and the final deployment models. This archive contains the following additional reproducibility artefacts. Associated repositories Computational study (code, trained models and reproducibility scripts)https://github.com/wells-wood-research/MHC_peptide_binding_model User-facing MHChron softwarehttps://github.com/wells-wood-research/MHChron PublicationDOI: add paper DOI once available Contents 1. Datasets This archive contains the datasets used throughout MHChron model development, including: the complete curated raw peptide–MHC dataset (~10 GB) assembled from the publicly available sources described in the accompanying manuscript, including experimentally determined peptide–MHC binding data and generated decoy peptide–MHC pairs; balanced subsampled datasets used during model development. These datasets form the starting point for the preprocessing, feature generation and model training pipeline implemented in the accompanying GitHub repository. 2. Cross-validation datasets Prepared datasets for all five cross-validation folds used during model development are provided for each evaluation regime: exploratory split; stratified split; leave-one-peptide-cluster-out (LOPCO) split; leave-one-allele-cluster-out (LOACO) split. For each fold, the archive includes the corresponding metadata describing the training, validation and test partitions. 3. External baseline comparison datasets The accompanying GitHub repository contains the scripts used to prepare inputs, execute external prediction tools and analyse the resulting predictions. This archive contains the generated input datasets required by the external predictors (for example, FASTA files and other tool-specific input formats where applicable), together with the exact subsets of the stratified test set on which each external method was evaluated. These files allow the published benchmarking experiments to be reproduced without regenerating the intermediate inputs. 4. Temporally held-out evaluation dataset This directory contains the temporally held-out experimental dataset used to assess model generalisation beyond the development data, as described in the accompanying manuscript. The dataset comprises experimentally determined peptide–MHC binding measurements deposited in the IEDB database between 16 July 2025 and 17 December 2025, which were unavailable during model development. Both the complete dataset and the balanced subsampled version used for evaluation are provided. 5. Data leakage analysis This directory contains the outputs of the data leakage analyses reported in the accompanying manuscript. Two versions are provided: analysis performed using the precomputed peptide and allele cluster assignments; analysis performed by recomputing the cluster assignments from scratch prior to analysis. These files demonstrate that the dataset splitting procedures satisfy the intended separation criteria and show no evidence of data leakage between the training, validation and test sets. 6. Deployment models This directory contains the final deployment models described in the accompanying manuscript. The deployment package includes: the MHChron-light model (.pt); the MHChron-edgy model (.pt); the fitted ESM feature scaler used by both deployment models. These models were trained on the complete development dataset, with a randomly selected 7% subset of the original stratified test set reserved for validation during training, as described in the manuscript. They correspond to the models distributed with the user-facing MHChron prediction software. 7. Cross-validation results The archive includes an Excel workbook containing the performance metrics obtained during model development across the different evaluation regimes. Results are reported separately for each of the five cross-validation folds for the exploratory, stratified, leave-one-peptide-cluster-out (LOPCO) and leave-one-allele-cluster-out (LOACO) experiments, providing the underlying data used to calculate the summary statistics reported in the accompanying publication. Not included This archive does not include the intermediate TFold-generated structures, AlphaFold pair representations, distograms, residue–residue distance maps or ESM embedding files generated during preprocessing. These intermediate artefacts require substantial storage and computational resources but can be regenerated from the raw datasets using the workflow provided in the accompanying GitHub repository. The complete source code, processed datasets required for routine use, deployment models, trained model weights and benchmarking scripts are distributed through the accompanying GitHub repository. Citation If you use these data or the accompanying software in your work, please cite the associated MHChron publication.