Quantum Companies Help Develop Hybrid AI for Immune-Targeting Peptides

Insider Brief

• Researchers developed a hybrid quantum-classical AI system that designed immune-targeting peptides, with quantum-generated inputs improving peptide discovery for understudied HLA variants and producing laboratory-validated binders.
• The quantum-enhanced model generated more predicted strong-binding peptides than conventional approaches, with the largest improvements occurring for HLA variants that have limited training data.
• The researchers caution that the study is a preprint and does not demonstrate quantum advantage, but suggests quantum-generated probability distributions could become a useful component of future AI-driven vaccine and immunotherapy design.

A hybrid quantum-classical artificial intelligence system generated short protein fragments that attached to immune-system proteins, with the largest gains appearing for genetic variants that are poorly represented in existing data, according to a preprint posted on bioRxiv. The findings suggest quantum-generated patterns could help AI systems search more effectively for vaccine and immunotherapy candidates.

The study, led by researchers from the Technical University of Denmark, ORCA Computing, Sparrow Quantum and several collaborating institutions, combined a photonic quantum processor with a conventional generative AI model. The system was designed to create peptides, which are short chains of amino acids, that can be displayed on the surface of cells for inspection by the immune system.

These displayed peptides can help T cells identify infected or cancerous cells. Finding peptides that attach securely to the proteins responsible for displaying them is considered an important early step in developing some cancer vaccines and T-cell therapies.

The researchers report that the project is, to their knowledge, the first complete hybrid quantum-classical system to design this type of peptide and then test selected results in the laboratory.

Unlike many quantum computing studies that remain entirely computational, the team synthesized some of the proposed peptides and confirmed that many formed stable complexes with their intended immune-system targets.

A Difficult Search Problem

Peptide design presents a vast search problem because each peptide is built from amino acids arranged in a particular order. Even for a peptide only nine amino acids long, the number of possible combinations is enormous. Only a small share will attach well to a given immune-system protein.

Those proteins are known as major histocompatibility complex class I proteins, or MHC-I. In humans, they are also described by their human leukocyte antigen, or HLA, types.

HLA types vary greatly among people with some that are common and supported by large amounts of experimental data. Others are rare or have unusual binding features, leaving AI systems with fewer examples from which to learn.

That uneven coverage is a significant factor for the development of personalized medicine. A vaccine or therapy designed around common HLA types may not work as well for people whose immune systems carry less common variants.

The researchers used this weakness as a test of whether quantum computing could improve how the AI system searched for promising peptide sequences.

They did not ask the quantum processor to run the entire AI model. Instead, they used it to generate the starting patterns that guide the model’s search.

Generative AI systems usually begin with random numerical inputs. Those inputs act as a seed from which the model creates a new image, molecule or biological sequence. Most systems use simple forms of random noise.

One way to think of it is to imagine that Generative AI starts with a random “starting ingredient,” much like a chef beginning with a mystery basket of ingredients. Different starting ingredients lead to different dishes, even though the same cooking method is used.

In this study, the team replaced that standard random input with patterns produced by a photonic quantum computer. Because photons can interfere with one another, the device produces complex relationships among its outputs rather than a collection of unrelated random values.

The researchers tested whether those more structured patterns could lead the AI model toward useful parts of the peptide search space that standard inputs might miss.

Gains Concentrated Among Difficult Targets

The team trained the system on about 106,000 previously observed peptide-HLA pairings, representing roughly 77,000 unique peptides and 126 HLA types.

The model was then asked to generate 1,000 candidate peptides for each of 131 HLA variants. A separate prediction program estimated how likely each peptide was to attach strongly to its assigned HLA protein.

Models supplied with quantum-generated starting patterns produced more likely binders than models using two conventional forms of random input.

The overall improvement was limited, but the gains were concentrated among HLA variants for which the conventional models performed poorly.

The simulated quantum approach produced an average of about 10.6 additional likely binders per 1,000 generated peptides compared with the standard Gaussian approach — or, a conventional random-number method based on a bell curve. The real photonic quantum processor produced about 6.3 additional likely binders per 1,000.

The quantum-based method outperformed the main conventional comparison on 63% of the HLA variants studied.

The researchers then examined three less-studied variants in greater detail. These were HLA-A31:01, HLA-A68:01 and HLA-B*37:01.

For HLA-A31:01 and HLA-A68:01, the models using quantum-generated inputs consistently produced more promising candidates during training. For HLA-A*68:01, the quantum-based models generated roughly twice as many likely binders as the conventional model.

Results for HLA-B*37:01 were less consistent. That variant has an unusual binding pattern and is poorly represented in existing datasets, making it especially difficult for prediction systems.

The team also examined whether the quantum-based model was repeating a narrow set of familiar peptide patterns and found that it actually generated a broader range of sequences.

The quantum-generated inputs led the model to explore a wider range of amino-acid combinations in parts of the peptide that can tolerate variation. At the same time, the model preserved the key amino acids needed to hold the peptide in place.

This is important because greater variety is useful only when it does not come at the expense of the features required for binding.

Laboratory Confirmation

Computer predictions alone are not enough, particularly for rare HLA variants where existing prediction tools may be less reliable.

The researchers therefore selected 20 of the highest-ranked peptides for each of the three HLA types and had them synthesized.

They then tested whether the peptides could form stable complexes with their assigned HLA proteins. A stable complex indicates that the peptide can remain attached long enough to be displayed to immune cells.

For HLA-A31:01 and HLA-A68:01, all 20 peptides tested formed stable complexes in the laboratory.

The HLA-B*37:01 results were more mixed with many peptides attaching successfully, while several did not.

The researchers said the uneven result likely reflects the difficulty of predicting binding for that HLA type rather than a weakness unique to the quantum method.

The strongest-performing peptide tested for each of the three HLA variants came from the quantum-based model, although the average laboratory results for the best quantum-generated and conventionally generated peptides were similar.

The experiments showed that the system was not simply producing sequences that looked promising to a computer model. At least some of its designs also worked in a biological test.

A Hybrid Role

The study points to a possible near-term role for quantum processors as supporting components within larger classical AI systems.

The photonic device did not train the neural network or calculate peptide binding. It supplied structured starting data while conventional graphics processors handled the main AI workload.

That approach may be more realistic than attempting to move an entire drug-discovery or biological-design program onto a quantum computer.

The processor used 32 optical modes and performed a method called Gaussian boson sampling. In simple terms, the device sent quantum states of light through an optical network and measured where the photons emerged.

Those measurement patterns became the starting inputs for the peptide generator.

The researchers also compared outputs from the physical device with computer simulations of the same quantum process. Their results were similar, suggesting that hardware imperfections did not erase the effect.

Limitations

The researchers cautioned that the work does not demonstrate quantum advantage.

For example, the system was small enough to be simulated on a conventional computer, even though those simulations took substantially longer than collecting samples from the physical quantum device.

It also remains possible that a more complex conventional probability distribution could produce similar results. The study compared the quantum method mainly with two common but relatively simple classical alternatives.

The choice of AI model presents another limitation. The researchers used a generative adversarial network, or GAN, because it allowed them to change the starting distribution without changing the rest of the system. That made the experiment easier to interpret, but GANs are not the newest or most capable tools available for peptide design.

Most important, binding to an HLA protein is only the first step toward producing an effective immune response.

A peptide that attaches securely may still fail to activate T cells. Immune activity also depends on whether the peptide is produced inside cells, whether it reaches the cell surface and whether suitable T cells exist to recognize it.

The laboratory tests measured the stability of the peptide-HLA complexes. They did not show that the peptides could prevent disease, kill cancer cells or work as vaccines.

Next Steps

The researchers suggest testing quantum-generated starting patterns with newer AI architectures, including systems designed specifically for protein and peptide structures.

Future models could also evaluate more than binding strength. They might account for how peptides are processed inside cells, how they change shape when attached to HLA proteins and whether they are likely to trigger a T-cell response.

Another possible next step would be to use experimentally confirmed peptides to improve the training data for rare HLA variants.

Such a system could create candidates, test them in the laboratory and then feed successful results back into the model. Over time, that process could reduce the data gap between common and uncommon immune-system variants.

For a deeper, more technical dive, please review the paper on bioRxiv. It’s important to note that bioRxiv is a pre-print server, which allows researchers to receive quick feedback on their work. However, it is not — nor is this article, itself — official peer-review publications. Peer-review is an important step in the scientific process to verify results.