Quantum-Classical GAN for MHC Class I Peptide Design

Peptides that bind major histocompatibility complex (MHC) class I molecules are the antigens that T cells survey to distinguish healthy cells from infected or cancerous ones, making their design central to vaccine development and T-cell therapies. Computational design remains difficult for the many human leukocyte antigen (HLA) alleles that are poorly represented in training data. This model is a hybrid quantum-classical conditional generative adversarial network (GAN) for de novo design of MHC class I-binding peptides, developed by the Technical University of Denmark with quantum-hardware partners ORCA Computing and Sparrow Quantum and computational collaborators at the MRC Laboratory of Molecular Biology, the Poznań Supercomputing and Networking Center, and Poznań University of Technology.

Its distinguishing idea is where the generator's randomness comes from. Instead of seeding the GAN with a conventional Gaussian noise prior, the system draws latent vectors from a real photonic quantum processor running Gaussian boson sampling. The intent is to inject structured, strongly correlated prior distributions that are difficult to reproduce classically, shaping the generator toward more diverse and productive regions of peptide space while the rest of the architecture is held fixed.

The work is a controlled test of whether a non-classical prior helps a generative biology pipeline, rather than an attempt to replace classical designers. The authors deliberately use a conditional GAN that is not state of the art precisely so that any difference in output can be attributed to the prior, and they position the result as an early bridge between quantum hardware and immunopeptidomics.

#Key Features

• Quantum-seeded latent space: Generator noise is sampled from a photonic quantum processor with 32 optical modes via Gaussian boson sampling, replacing the usual classical Gaussian prior.
• Allele-conditional generation: The GAN is conditioned on HLA identity, generating candidate 9-mer peptides for a specified MHC class I molecule.
• Gains for underrepresented alleles: The largest improvements appear for understudied alleles such as HLA-A*31:01, HLA-A*68:01, and HLA-B*37:01.
• Experimental validation: Top-ranked designs were synthesized and assessed by peptide-MHC stability ELISA rather than left as in-silico predictions.
• Honest scope: The authors state that no quantum advantage is demonstrated, since the system is small enough to simulate on classical hardware.

#Technical Details

The generator was trained on roughly 77,000 unique validated 9-mer peptide ligands (about 106,000 peptide-HLA pairings) drawn from the Immune Epitope Database and labeled with NetMHCpan-4.1, spanning 126 MHC class I molecules; generation was targeted across 131 HLA alleles. Using a quantum-sampled prior, simulations produced approximately 10.6 additional predicted binders per 1,000 generated peptides relative to a Gaussian baseline, while the physical photonic device yielded about 6.3 additional binders, and the quantum-seeded approach outperformed the classical baseline on 63% of evaluated HLA variants. For in vitro validation, the 20 highest-ranked peptides per allele were synthesized: for HLA-A*31:01 and HLA-A*68:01 all 20 formed stable peptide-MHC complexes, whereas HLA-B*37:01 gave mixed results.

#Applications

The method addresses immunogen and epitope design for vaccines, neoantigen discovery, and T-cell therapies, where finding binders for less-studied HLA alleles is a persistent bottleneck. Alleles such as HLA-A*31:01, which is associated with severe drug hypersensitivity reactions, illustrate the clinical relevance of extending reliable design to underrepresented genetic backgrounds. Immunologists, vaccine developers, and computational protein designers are the primary beneficiaries, and the peptide-MHC stability assays provide a wet-lab readout that connects generated sequences to measurable binding.

#Impact

The study offers a concrete demonstration that photonic quantum hardware can be integrated into a generative peptide-design pipeline and produce candidates that hold up under experimental binding assays. Its contribution is methodological: a clean comparison isolating the effect of a quantum-sampled prior within an otherwise standard conditional GAN. The authors are explicit about the limits, noting that the system is classically simulable at this scale so no quantum advantage is claimed, that a richer classical prior might achieve similar gains, and that binding affinity alone does not guarantee immunogenicity. The work is a bioRxiv preprint without a released code or weights repository, positioning it as an early proof of concept at the interface of quantum computing and immunopeptidomics.