Hybrid quantum-classical de novo design of MHC-binding peptides

Technical Summary: Hybrid Quantum-Classical De Novo Design of MHC-Binding Peptides

Problem Statement
The design of therapeutic peptides, specifically those that bind to Major Histocompatibility Complex (MHC) class I molecules, relies on navigating vast, high-dimensional sequence spaces. While deep generative models have emerged as a leading approach for this task, they face significant challenges when training data is sparse, particularly for understudied or rare Human Leukocyte Antigen (HLA) alleles. Standard generative models, such as Generative Adversarial Networks (GANs), typically utilize simple, factorized classical distributions (e.g., isotropic Gaussian) as priors to seed the generation process. The authors posit that these uncorrelated priors may limit the model's ability to explore complex, structured regions of the sequence space where biological constraints (such as specific anchor residues) coexist with high variability in non-anchor positions. The central question is whether non-classical, structured prior distributions sampled from quantum processors can serve as a more effective inductive bias for de novo peptide design, particularly in data-poor regimes.

Methodology
The authors developed an end-to-end hybrid quantum-classical pipeline to test this hypothesis:

1. Dataset Construction: A curated dataset of 105,970 experimentally validated peptide-HLA pairs (approximately 77,000 unique 9-mer sequences) was assembled from the Immune Epitope Database, covering 126 distinct HLA alleles.
2. Model Architecture: A conditional GAN was employed. The generator maps latent vectors and HLA allele labels to peptide sequences, while the discriminator distinguishes between generated and real peptides. The architecture was kept simple and consistent across all experiments to isolate the effect of the prior distribution.
3. Prior Distributions: The study compared four prior distributions of identical dimension (32):
   • Classical Baselines: Factorized Gaussian and Bernoulli distributions.
   • Quantum Distributions: A simulated Gaussian Boson Sampler (GBS) and a real distribution sampled from a 32-mode photonic quantum processor (using time-bin interferometry and superconducting nanowire detectors).
4. Training and Evaluation: Models were trained on the pMHC dataset. For 131 HLA alleles (including 5 not in the training set), 1,000 peptides were generated per allele. Binding affinity was predicted using NetMHCpan 4.2, with "strong binders" defined by a percentile rank threshold of 0.5%.
5. In Vitro Validation: Three understudied alleles (HLA-A68:01, HLA-B37:01, and HLA-A*31:01) were selected for experimental validation. The top 20 predicted binders per allele (generated using the real quantum prior) were synthesized and tested using peptide-MHC stability ELISAs following UV-mediated ligand exchange.

Key Results

• Increased Binder Yield: Models conditioned on quantum-derived priors (both simulated and real) produced a higher number of predicted strong binders compared to classical Gaussian and Bernoulli baselines across the 131 alleles.
• Performance in Data-Sparse Regimes: The improvement was most pronounced for rare alleles where classical baselines performed poorly. The quantum prior outperformed the Gaussian prior on 63% of alleles, with the largest relative gains observed in alleles with low training set frequency.
• Sequence Diversity: Analysis of Shannon entropy revealed that quantum priors facilitated broader exploration of non-anchor positions while maintaining strict specificity at anchor positions. This suggests the quantum prior enables the model to sample a wider range of allele-compatible sequences without losing the critical structural motifs required for binding.
• Experimental Confirmation: In vitro ELISA assays confirmed that peptides generated using the real photonic quantum processor successfully stabilized peptide-MHC class I complexes for all three tested alleles. Notably, for HLA-B*37:01—an allele with an atypical acidic anchor sparsely represented in training data—the quantum-prior model produced strong binders, whereas classical predictors are known to be less reliable for such targets.

Significance and Claims
The paper claims to present the first end-to-end hybrid quantum-classical pipeline for the de novo design of MHC-binding peptides. The authors do not claim a demonstration of "quantum advantage" in the computational complexity sense, noting that the system sizes used are classically simulable. Instead, the significance lies in establishing structured, hardware-realizable non-classical priors as a useful inductive bias for generative peptide design.

The work demonstrates that the structure of the prior distribution fundamentally shapes how generative models explore sequence space. By leveraging the high-dimensional, correlated probability distributions inherent to photonic quantum processors, the method improves the yield of functional designs, particularly for alleles that are biologically constrained and underrepresented in existing datasets. This approach offers a practical pathway for enhancing personalized immunotherapies and vaccines, where the ability to design binders for rare or patient-specific HLA alleles is critical. The results suggest that quantum-derived priors can effectively address the limitations of current training data coverage in immunoinformatics.