Discovery and Validation of Physicochemical Determinants for Pan-Allele Peptide-MHC Class I Modelling Across Binding Affinity, Antigen Presentation, and Immunogenicity

Problem. Personalised cancer-vaccine design requires predicting which of a patient’s mutated peptides will be presented to the immune system. The decision rests on three sequential biological filters: a peptide must bind a major histocompatibility complex (MHC) class I molecule (binding affinity, BA), the complex must be processed and displayed at the cell surface (eluted-ligand presentation, EL), and the displayed complex must be recognised by a CD8+ T-cell receptor (immunogenicity, IM). Computational prediction of the first two filters has been the rate-limiting step for two decades, and NetMHCpan-4.1 remains the field’s reference predictor.

Gaps and research questions. Reviewing that lineage identifies three gaps. Existing pan-allele predictors transfer poorly to the long tail of low-frequency human leukocyte antigen (HLA) alleles; they under-fit non-canonical peptide lengths; and they consume the 34-residue MHC pseudo-sequence — the structurally derived map of which binding-groove positions contact the peptide —only as a flat embedding or an allele identifier, never as a structural target that individual peptide positions can address. A fourth, methodological concern is that headline comparisons in this literature are often reported without confidence intervals, from single folds, or without decontaminating the benchmark against the dominant training corpus. These motivate four research questions: whether position-addressable cross-attention over the pseudo-sequence improves prediction (RQ1); whether contrastive regularisation grouped by structural similarity rather than allele identity improves rare-allele transfer (RQ2); whether one pan-length model suffices for non-canonical lengths (RQ3); and whether, under a declared evidential contract, the resulting predictor outperforms NetMHCpan-4.1— and where it does not (RQ4).

Approach. This thesis develops DeepNeo-CL, a peptide–MHC-class-I predictor built on a frozen ESM2-35M protein language-model trunk extended with position-resolved physicochemical features, peptide-conditioned cross-attention over the pseudo-sequence, joint BA and EL output heads, long-tail imbalance correction, and a supervised contrastive refinement grouped by pseudo-sequence cluster. A compact immunogenicity head reuses the same trunk, so one forward pass returns BA, EL and IM scores. Training used the official NetMHCpan-4.1 five-fold corpus (207,029 BA pairs and approximately 33 M EL pairs), and every Stage 1 result reported here is a five-fold ensemble. The Stage 2 immunogenicity head is trained and reported separately, as a three-seed ensemble over the frozen Stage 1 trunk.

Experiments. Evaluation used three independent held-out benchmarks — TransPHLA (n = 161,652), an IEDB BA snapshot decontaminated at peptide–HLA pair level against the training corpus (n = 1,365), and two years of the official IEDB weekly auto_bench series — plus BigMHC’s held-out immunogenicity test set for the Stage 2 head. Comparators were NetMHCpan-4.1 and MHCflurry-2.0, extended to BigMHC, MHCfovea, TransPHLA and ImmuneApp on identical post-cutoff rows. A declared three-condition evidential contract — bootstrap confidence intervals, paired DeLong testing, and per-allele Wilcoxon agreement — is the standard against which the comparisons are reported, and the thesis states plainly what that standard is worth: it is not a pre-registration, and it was not applied in full. The DeLong condition is defined for the area under the receiver-operating-characteristic curve (AUROC) only, so no comparison of areas under the precision–recall curve (AUPR) lies within its reach; the paired bootstrap interval the first condition asks for was never computed, so no verdict satisfies all three. A cumulative ablation isolated each component’s contribution.

Results. The outcome is two-sided. On TransPHLA the ensemble exceeds NetMHCpan-4.1 on three of four metrics and ties on the fourth: BA AUROC/AUPR reach 0.9694/0.9705 through a label-free BA+EL rank fusion (DeLong z = 22.9), EL AUROC reaches 0.9697 (p = 1.8 × 10−8), and EL AUPR is tied (0.9714 vs 0.9716). On IEDB the two predictors are not distinguished on within-allele ranking across two years of weekly data (Wilcoxon p = 0.65 — a failure to detect a difference, not a demonstration of equivalence), while on the static snapshot DeepNeo-CL trails both on pooled cross-allele ranking and within allele, and percentile-rank recalibration does not repair that deficit. MHCflurry-2.0 is tied on IEDB. The ablation shows no single component worth more than +0.005 AUROC. Taken in pairs the contributions are larger but still modest: the pseudo-sequence family — concatenation together with cross-attention — contributes +0.007 after the physicochemical rung, the contrastive refinement is within noise pooled yet worth +0.022 AUPR on rare alleles, and the fusion lifts BA AUROC from the BA head’s 0.9546. An exploratory structural rescorer was withdrawn when its advantage proved transductive. The immunogenicity head reaches AUROC 0.5735, a statistical tie with BigMHC’s published predictor (p = 0.70), with no dedicated immunogenicity predictor separating from it after multiple-comparison correction.

What this means. The answer to RQ1 is a qualified yes — the mechanism works, but its isolated effect is small and it costs a fixed ≈ 10 per cent of the forward pass, so it is a modest trade rather than a free one. RQ2 is answered yes on the subgroup it targets, and the structural grouping rather than contrastive learning as such is what carries it. RQ3 is answered only partly, since the benchmark covers 8–11mers and cannot speak to the longer peptides that motivated the gap. RQ4 is answered three ways: yes on TransPHLA, tied on IEDB within-allele, and no on pooled IEDB. The defensible conclusion is therefore narrow: several biologically motivated refinements can together produce a measurable improvement on an independent benchmark, with the margins small, one benchmark unfavourable, and the evidential standard declared in advance of the results but neither pre-registered nor fully applied. The within-allele tie and the pooled deficit are measured on different corpora — two years of auto_bench and a single 1,365-row snapshot respectively —so together they bound what can be claimed rather than isolating a single cause; percentile-rank recalibration does not repair the pooled deficit, so it is not a score-scale artefact. The thesis makes no claim on TCR specificity, class II MHC, or post-translationally modified peptides, and advances no clinical claim. Alongside the thesis a reproducibility vault is released — code, ensemble checkpoints with integrity hashes, raw prediction arrays, statistical logs, figures, and a CPU-only validator — from which the base-ensemble results can be re-derived. The thesis also records which results lie outside it, the fused headline figure among them, and names the artefacts that produce those, so that every reported number can be traced to its source even where that source is not the vault.