A3 - MISTIC-MHC: An Alignment-free Allele-specific Approach for Robust Peptide-HLA Binding Prediction

Introduction: : Accurate prediction of peptide binding to human leukocyte antigen (HLA) molecules is central to vaccine design and immunotherapy. However, high polymorphism across >39,000 Class I and II alleles and variable peptide lengths, especially in Class II, complicate prediction. While deep learning tools like NetMHCpan and ESM2-NN offer strong performance, they often lack interpretability and struggle with limited or imbalanced datasets. We present MISTIC-MHC, a feature selection and classification framework combining Fourier-transformed sequence encoding with Support Vector Machines (SVMs). Peptides are encoded using biologically-informed oscillation patterns derived from BLOSUM matrices, enabling standardized, interpretable inputs. MISTIC ranks features using two complementary strategies: Feature Importance, measuring global model sensitivity, and Feature Contribution, quantifying feature-specific influence on predictions. These are integrated into a consensus ranking that guides feature selection via greedy backward elimination. MISTIC enhances both performance and interpretability across allele-specific models for Class I and II HLA molecules.

Materials and

Methods: : We curated quantitative HLA-peptide binding data from IEDB, resulting in ~72K Class I and ~68K Class II entries across 145 alleles. Peptides were numerically encoded via FFTpro, which applies Fast Fourier Transform to BLOSUM-encoded amino acid sequences, generating 120 features per peptide. Nonlinear SVM classifiers with RBF kernels were trained per allele and optimized through stratified cross-validation and grid search. MISTIC’s feature selection employs two complementary evaluations: Feature Importance (based on SVM objective sensitivity) and Feature Contribution (based on decision boundary perturbation). These metrics are combined into a consensus rank, and feature subsets are iteratively reduced using greedy backward selection. Model performance was assessed using AUC and F1 score, with a balanced consensus metric. We benchmarked MISTIC against ESM2-NN and NetMHCpan using a custom blind test set to evaluate generalization across alleles and dataset types.

Results, Conclusions, and Discussions:: MISTIC achieved strong predictive performance across Class I and II alleles. Optimal models were derived by balancing Feature Importance and Contribution, with higher weight(w) on Contribution improving results, particularly on small or imbalanced datasets. AUC and F1 scores remained high even with reduced feature sets, confirming MISTIC’s efficiency and interpretability (Figure 1). We introduced an “enrichment score” to identify the best feature ranking strategy per allele based on the area under the AUC-vs-feature-count curve.

On blind test sets, MISTIC outperformed ESM2-NN for most alleles, especially those with limited data (Figure 2). Compared to NetMHCpan, MISTIC performed competitively, often outperforming on HLA-C and HLA-DR alleles (Figure 3). While NetMHCpan had an edge on HLA-A alleles with large, balanced datasets, MISTIC showed better generalizability without prior exposure to test data. These results suggest MISTIC is especially well-suited to data-limited or interpretability-focused applications.

MISTIC’s biologically grounded features and transparent selection process provide insight into molecular drivers of peptide binding, offering practical advantages for epitope discovery and immunotherapy design. Its open-source availability facilitates integration into broader bioinformatics pipelines. Overall, MISTIC-MHC is a robust and interpretable solution for allele-specific peptide binding prediction across diverse HLA types.

Acknowledgements and/or References (Optional)::