Structure-Guided Benchmark of pMHC-I Peptides

• Structure-guided benchmarking of pMHC-I peptides is a method that uses three-dimensional structural data to assess binding and immunogenicity.
• It employs techniques like stability calculations, contact map analysis, and diffusion model generation to curate unbiased peptide libraries.
• This approach exposes limitations in conventional predictors and informs improvements in immunotherapy and vaccine design.

Structure-guided benchmarking of peptides bound to MHC class I molecules (pMHC-I) refers to computational and experimental frameworks that leverage three-dimensional structural information—including peptide conformational stability, residue interaction networks, and proximity to MHC binding pockets—to create more rigorous, biophysically faithful standards for evaluating peptide immunogenicity and MHC binding affinity. These benchmarks address critical shortcomings in traditional sequence-based and MS assay–derived approaches, particularly by revealing aspects of immunogenicity and binding not captured by high-throughput or shallow predictive models.

1. Principles of Structure-Guided Benchmarking

The core rationale for structure-guided benchmarking in the pMHC-I context is that peptide immunogenicity and binding are fundamentally determined by the physical compatibility between the peptide and the MHC groove, as well as by the resultant conformational integrity required for T cell recognition (1004.3175). Classical benchmarks often emphasize affinity values measured biochemically or deduced from mass spectrometry, but such approaches may neglect the spectrum of physically feasible, structurally stable peptides not covered by experimental biases or sequence diversity limitations (Mares et al., 11 Jul 2025).

In structure-guided approaches, atomic-level features—such as residue–residue contact patterns, solvent-accessible surface area, and energetic stability computed over static or modeled complexes—directly inform the selection and evaluation of benchmark peptides. This allows benchmarking resources to prioritize peptides exhibiting canonical anchor residue preferences, favorable hot-spot interactions, or peptide backbone geometries compatible with the MHC class I groove, independently of training set frequency or assay representation.

2. Methodologies for Structure-Based Benchmark Creation

A wide array of computational techniques has been described for generating structure-informed peptide benchmarks:

• Stability-Based Filtering: Peptide stability is computed using energy functions such as the AMBER force field, where peptide “self-energy” (the intrinsic energy of the isolated peptide) is compared to “full energy” (including interactions with the source protein). The energy difference $$E_{\text{difference}} = E_{\text{full}} - E_{\text{self}}$$ functions as a proxy for how robust a peptide is in isolation—potentially predicting its likelihood to maintain a fold compatible with MHC binding (1004.3175).
• Contact Map–Guided Metrics: Crystal structures are processed to build contact maps, often using all heavy-atom Euclidean distances. Contacts at or below a distance threshold (e.g., $d \leq 3.5$ Å) are flagged as hot-spots, reflecting anchor residues or energetically pivotal interactions (Mares et al., 11 Jul 2025). Structure-aware diffusion models use these contact maps as constraints during peptide library generation.
• Diffusion Model Generation: Geometry-aware generative models such as RFdiffusion introduce noise to peptide backbones and iteratively “denoise” them while conditioning on the MHC groove and fixed hot-spot residues. This process yields libraries of sequences structurally optimized for MHC compatibility. Subsequently, side-chain packing (e.g., via ProteinMPNN) and structure prediction (e.g., via AlphaFold2-Multimer) further refine and filter candidates, retaining only those with high predicted integrity (pLDDT > 0.8) (Mares et al., 11 Jul 2025).
• Energy-Based Design and Sampling: Physics-guided machine learning frameworks such as HERMES predict amino acid probabilities at each peptide position based on the 10 Å local environment. Peptide libraries are then generated either from fixed backbone templates or through iterative Markov Chain Monte Carlo (MCMC) packing/sampling protocols (Visani et al., 1 Mar 2025).

3. Benchmark Datasets and Resource Characteristics

Recent advances have produced several structure-guided or geometry-aware benchmark datasets for pMHC-I peptides:

• Independence from Standard Assays: Unlike traditional collections—often restricted by bias toward abundant/self peptides—structure-guided benchmarks can be synthesized to be “orthogonal” to experimentally characterized sets, thus offering a more comprehensive view of the peptide–MHC landscape (Mares et al., 11 Jul 2025).
• Allele Breadth and Residue Diversity: Resources such as the RFdiffusion-generated benchmark span up to twenty high-priority HLA alleles and include 9–11mer peptides. Despite being independent of training distributions, these libraries robustly recover canonical anchor preferences and display higher per-residue diversity than seen in MS-derived datasets (Mares et al., 11 Jul 2025).
• Availability: Code, diffusion pipelines, and benchmarks are distributed openly, e.g., https://github.com/sermare/struct-mhc-dev (Mares et al., 11 Jul 2025).

The following table compares key aspects of recent structure-guided benchmarks:

Benchmark Type	Dataset Bias-Free	Structure Conditioned	Residue Diversity
MS assay–derived (canonical)	No	No	Limited
Diffusion-based, structure-guided (Mares et al., 11 Jul 2025)	Yes	Yes	High

4. Evaluation of Predictive Models Using Structure-Guided Benchmarks

Benchmarking predictive performance on structure-guided peptide sets illuminates substantial limitations in state-of-the-art sequence-based approaches. Key findings include:

• AUROC Degradation: While predictors such as MHC-Flurry, NetMHCSpan, and MixMHCpred achieve AUROC values between 0.66 and 0.81 on experimental peptide datasets, their performance on geometry-optimized, structure-guided benchmarks deteriorates dramatically (AUROC 0.06–0.22), suggesting lack of generalization to structurally fit, novel epitopes (Mares et al., 11 Jul 2025).
• Anchor Recovery and Motif Generalization: Structure-guided benchmarks still reproduce hydrophobic anchor residue enrichment at positions P2 and Pω, confirming that the generative process respects fundamental MHC-I binding motifs, but the synthetic diversity of backbone and side-chain combinations exposes weak spots in training-dependent predictors.

This suggests that model validation on structure-informed benchmark sets is essential to reveal allele-specific limitations and to guide architectural or training improvements in peptide–MHC predictive frameworks.

5. Integration of Structural Insights in Peptide Design and Immunogenicity Prediction

Structure-guided benchmarks provide the foundation for designing new immunogens and probing the boundaries of TCR recognition:

• Energetic and Entropic Modeling: Physical energy computations, when combined with measures such as peptide entropy from position weight matrices, allow quantification of TCR specificity and cross-reactivity. For example, the median peptide entropy recognized by a given TCR–MHC pair is approximately 10 bits (implying ∼$10^3$ peptides recognized), in contrast to the $\sim31$ bits of peptide entropy available to MHC-I alone ($\sim2 × 10^9$ distinct peptides) (Visani et al., 1 Mar 2025).
• Neoantigen Prioritization: Models that weight residue contacts by structural recurrence (i.e., contact maps) inform which mutations are likely to be immunodominant. Mutations occurring at high-contact positions yield larger perturbations in binding energy—and thus, are more likely to be recognized by T cells (Chau et al., 2021).
• Experimental Validation: Newly designed peptides, informed by structure, have shown up to 50% experimental success in T cell activation assays despite bearing multiple substitutions from the native epitope (Visani et al., 1 Mar 2025).

6. Implications and Future Directions

Structure-guided benchmarking represents a paradigm shift in the evaluation and design of pMHC-I peptides:

• Model Development: Geometry-aware benchmarks expose deficiencies in current predictors, supporting the development of new architectures that incorporate structure—such as transformer models with cross-attention guided by residue proximity, or generative models calibrated on unbiased, structure-centric datasets.
• Immunotherapy and Vaccine Design: By enabling a less biased exploration of peptide space, structure-guided libraries can accelerate the discovery of unconventional but structurally viable epitopes, broadening the repertoire of targets for personalized vaccines and immunotherapies (Mares et al., 11 Jul 2025).
• Community Resources: Publicly available benchmark sets and codebases provide a foundation for reproducible research and foster cross-group comparisons of algorithmic advances.

A plausible implication is that as structure-guided benchmarks become more widely adopted, future predictive models and therapeutic pipelines will necessarily integrate three-dimensional structural analysis to achieve robust, quantitative, and generalizable performance in pMHC-I peptide discovery and ranking.