STRUMP-I

Predicting which short peptides a given MHC class I allele will present is the core filtering step in nearly every cancer neoantigen and T-cell vaccine pipeline. The tools that dominate practice — NetMHCpan, MHCflurry, HLAthena — read peptide and MHC sequences directly and are fast, but their accuracy tracks how much binding and eluted-ligand data exists for each allele. MHC-I is the most polymorphic locus in the human genome, and most of its thousands of alleles have few or no measured ligands, so pan-allele sequence models degrade exactly where clinical genotypes are most likely to fall.

STRUMP-I (STRUcture-based pMHC Prediction for class I) attacks that gap from the physics side. Rather than searching for a single minimum-energy binding pose — an assumption that sits badly with the known promiscuity of MHC-I grooves — it builds a homology model of each peptide-MHC complex, relaxes and repacks it, and then feeds the entire vector of force-field energy terms into a trained classifier as features. The classifier learns which combination of solvation, electrostatic, and packing terms distinguishes binders, without ever needing to have seen the allele. It was developed by researchers at Chonnam National University Medical School, Boston Children's Hospital, and Samsung Medical Center, with the Korean biotech firms Arontier and Geninus, and posted as a preprint in September 2025.

The result is a structure-based predictor that matches sequence-based tools on their own turf while clearly beating them on sparsely characterized alleles — closing a gap that AlphaFold-derived approaches such as AlphaFold-FineTune and MHC-Fine had narrowed but not erased.

#Key Features

• Energy terms as features, not as a score: Instead of collapsing FoldX energetics into one binding-affinity proxy, STRUMP-I passes the full set of decomposed terms to the classifier. The raw FoldX interaction energy alone reaches only 0.64 AUROC; the learned combination reaches 0.91.
• Robustness on underrepresented alleles: On test alleles with fewer than 500 training peptides, STRUMP-I holds a PRAUC of 0.989 and MCC of 0.881 while MHCflurry falls to 0.732 MCC and HLAthena to 0.500.
• Precision filter for existing pipelines: Because its false positives are largely disjoint from those of sequence-based tools, stacking STRUMP-I onto them lifts combined precision to 0.991 on the held-out IEDB+HLAthena set.
• Structural byproducts: Every prediction yields a relaxed, side-chain-optimized pMHC model, reusable for solvent-accessibility analysis and downstream TCR work.
• Distributed pretrained model: The repository ships an installable strump_i package with the trained classifier, scaler, and 826 curated IMGT templates, applied to new alleles without retraining.

#Technical Details

For a query peptide and allele, STRUMP-I selects a template from 826 IMGT-derived pMHC-I structures, matching peptide length first, then allele similarity by BLOSUM62 and peptide similarity by PAM30 over the 181-residue MHC-I G-domain. FoldX BuildModel installs the query sequence into the template; Tinker 8.9.5 relaxes the backbone under AMBER99sb with the GB-HPMF implicit solvent model to clear steric clashes; FoldX then optimizes side chains to convergence and AnalyzeComplex emits the energy terms. Those terms, plus template alignment scores, form the feature vector for a LightGBM 4.0.0 classifier tuned with Optuna over 10,000 rounds. Training used a merged IEDB and HLAthena mono-allelic eluted-ligand set of 288,269 peptide-MHC pairs (207,359 binders at a 500 nM cutoff, 81,110 non-binders) split 8:1:1. On the held-out test set of 38,302 binders and 9,061 non-binders across 115 alleles, STRUMP-I scores 0.984 PRAUC and 0.756 MCC, against 0.982/0.781 for MHCflurry, 0.986/0.778 for HLAthena, 0.974/0.603 for NetMHCpan, and 0.916/0.422 for AlphaFold-FineTune. SHAP analysis attributes importance to non-bonded interaction, solvation, and electrostatic terms — and, notably, to template alignment quality.

#Applications

The intended user is a neoantigen or vaccine pipeline that needs candidate peptides worth synthesizing. Sequence tools screen a patient proteome cheaply; STRUMP-I then rescores the shortlist to strip false positives. On an experimentally assayed set of neoantigen candidates from 33 patients with colorectal, hepatocellular, gastric, and melanoma tumors, it cut false positives from 38 to 28 and raised precision to 0.870 over the NetMHCpan/MHCflurry prefilter. It is also the more defensible choice when a patient's HLA type is rare enough that sequence-model training data is thin.

#Impact

STRUMP-I is a preprint awaiting peer review, and its scope is deliberately narrow: a fixed gradient-boosting head over engineered physics features, not a learned representation. Its costs are real: building and relaxing a structure per peptide is far slower than a sequence forward pass, and FoldX and Tinker must be installed separately. It is also weaker than sequence tools on well-represented alleles, trailing on the PRIME benchmark at 0.846 PRAUC against 0.953-0.988. The contribution is the demonstration that treating a force field's decomposed output as a learnable feature space, rather than as an energy to minimize, closes the accuracy gap that has kept structure-based pMHC prediction on the sidelines. Code and trained models are released under the MIT license.