Simulated Quantum Annealing Ranked Known Drug Pairs First for Diabetes and Rheumatoid Arthritis: Ramos, Coutinho and Magano in Quantum Machine Intelligence, 23 June 2026
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · September 24, 2026.
Three researchers from Porto and the German Aerospace Center have written drug-pair selection as an energy problem that a quantum annealer could solve, and tested it on diabetes, rheumatoid arthritis, asthma and brain tumors. In simulation, the lowest-energy answers were dominated by drug combinations that doctors already use, and one of the top new suggestions for rheumatoid arthritis matches a pairing that a 799-patient trial found to roughly double remission rates.
The paper, by Diogo Ramos, Bruno Coutinho and Duarte Magano, appeared open access in Quantum Machine Intelligence on 23 June 2026 as article 70 of volume 8. It builds on the Complementary Exposure idea from network medicine, turns it into an optimization problem, and samples answers with simulated quantum annealing, a classical computer program that imitates how quantum annealing hardware behaves. Its central test is how often validated drug combinations appear among the top-ranked answers when they make up less than a tenth of a percent of the candidates.
Combination therapy is standard across chronic disease and oncology, because two drugs together can allow lower doses and fewer side effects. The search space is the problem. With more than a thousand approved drugs and thousands of diseases, the number of possible pairs grows far faster than any laboratory or trial program can test them.
For a pharmacology team or a hospital formulary committee, the study offers a clear example of what quantum optimization can and cannot yet contribute to that search. The ranking method works in simulation on four diseases. No quantum computer was used, and the authors state plainly that their predictions are hypotheses for experts to examine.
How Ramos, Coutinho and Magano turned drug pairing into an energy problem
The starting point is a map of the human protein-protein interactome: 16,677 proteins connected by 243,603 physical interactions, each supported by direct experimental evidence, compiled by Feixiong Cheng, István Kovács and Albert-László Barabási. In their 2019 study in Nature Communications, Cheng and colleagues showed that effective drug pairs tend to follow one pattern on this map. Each drug's protein targets sit close to the disease's cluster of proteins, while the two drugs' targets stay apart from each other. They called this Complementary Exposure: both drugs hit the disease, each from a different side.
Ramos, Coutinho and Magano encode that pattern as a quadratic unconstrained binary optimization problem, the input format a quantum annealer accepts. Each candidate drug is a yes-or-no variable. A drug lowers the energy when its targets lie near the disease cluster, a pair raises it when the two drugs overlap too much, and a penalty term holds the answer to two or three drugs. The lowest-energy configurations are the proposed combinations.
Two tuning parameters set the balance between those terms. The authors calibrated them for each disease on a benchmark of known effective combinations, using candidate pools of nine or ten drugs, small enough to check every possible answer exactly. Then they widened each pool to 50 drugs by adding randomly selected drugs, leaving out any whose targets overlapped strongly with the known drugs. That gives 20,825 allowed combinations of two or three drugs per disease, and the authors asked whether the known combinations still rose to the top. This is a test of how well the method recovers combinations it was tuned on, inside a larger pool; the known combinations stay in both pools, so it is no held-out clinical validation.
Quantum pillar: computing. Technology readiness: TRL 2 of 9. The method exists as published software tested only in a classical simulation of quantum annealing on existing data, with no run on quantum hardware, no laboratory test of the new drug pairs and no patient involved.
What the simulation found for diabetes, rheumatoid arthritis, asthma and brain tumors
On the 50-drug pools, four of the five lowest-energy answers were validated combinations for diabetes mellitus, rheumatoid arthritis and brain neoplasms. At rank 10 the share was between 40 and 60 percent for all four diseases. Across 1,024 independent runs of the simulated annealer on the diabetes pool, about 40 percent returned a validated combination among the ten lowest-energy answers, in a space where validated pairs make up 0.08 percent of the candidates. The corresponding figures were 33 percent for rheumatoid arthritis, 23 percent for asthma and 17 percent for brain tumors.
The new suggestions are the more interesting part for clinicians. For rheumatoid arthritis, the model ranked adalimumab with methotrexate third, although that pairing was not in its benchmark. A double-blind trial led by Ferdinand Breedveld, published in Arthritis and Rheumatism in 2006, randomized 799 patients with early, aggressive rheumatoid arthritis and found that after two years 49 percent of patients on the combination were in remission, about twice the rate on either drug alone. For diabetes, the top new pair was metformin with irbesartan. A 2012 laboratory study by Ishibashi and colleagues in Pharmacological Research found that irbesartan added to metformin's protection of kidney tubular cells against diabetes-related injury in cell culture.
The authors report two weaker results with the same care. Several asthma suggestions featured arformoterol, a bronchodilator approved for chronic obstructive pulmonary disease, which may reflect the closeness of the two diseases on the protein map. The brain tumor suggestions repeatedly included vinorelbine, epirubicin, tamoxifen and capecitabine, drugs used mainly in breast cancer. That pattern points to shared cancer pathways in the network, and it would need specialist review before anyone read it as a brain tumor finding.
Why the classical annealer matched the quantum simulation on three of four diseases
The paper also compares its simulated quantum annealer with an ordinary classical annealer, both taken from D-Wave's Ocean software library and given identical schedules. The classical baseline matched or outperformed the quantum simulation for diabetes, rheumatoid arthritis and asthma. Brain tumors were the exception: at a pool of 14 drugs the quantum simulation found the lowest-energy answer 28 percent of the time against 19 percent for the classical method, an advantage that faded as the pool grew.
The authors connect that exception to theory, which predicts that quantum tunneling helps most when the energy landscape has tall, thin barriers, and the brain tumor problem had the most rugged energy surface of the four. They also warn that simulated quantum annealing is a classical imitation, and whether the effect would appear on real annealing hardware remains an open experimental question. This Monitor made the same point on 23 September about simulated quantum kernels that lost every matched test to a default classical classifier on brain MRI and breast ultrasound: a quantum method earns its place only against a strong classical method run under the same conditions.
What limits the result: small benchmarks, an incomplete protein map and no hardware run
The benchmark is small. For some diseases the calibration rests on as few as 11 known combinations among 165 possible answers, and the tuned parameters vary widely between diseases, from 3.79 for asthma to 34.38 for brain tumors. The authors say this may partly reflect overfitting, and they did not hold out part of the benchmark for a separate test because large validation sets do not yet exist.
The protein map is incomplete too. The interactome used here captures an estimated 25 to 35 percent of all true physical interactions in human cells, and well-studied proteins are overrepresented. Every network medicine method shares that limitation, and the authors cite robustness checks from earlier work that support the key distance measures.
Finally, the method ranks two- and three-drug combinations by their position on a network. A low energy score says nothing direct about dose, safety, drug interactions or effect in patients. The authors write that their outputs should be read as ranked hypotheses, and they have published their code and benchmark on GitHub so others can test them.
How Quentir Reads It: what pharmacology teams and hospital buyers should ask in 2026
For drug repurposing groups, the study shows a transparent way to shortlist combinations from network data, and the classical version of the same ranking already works today on ordinary computers. A quantum annealer adds value only if it finds good answers faster or more often on larger pools, and this paper leaves that question open.
For hospital and health-system buyers, a vendor offering quantum drug-combination discovery should be asked three concrete questions. Was the method run on quantum hardware or in simulation? How did it compare with a classical annealer under the same budget? Which of its suggestions have been tested in a laboratory or in patients?
For patients, nothing in current treatment changes. The rheumatoid arthritis pairing the model recovered has been in clinical use for many years, and the method recovered that established pairing although it was absent from its benchmark, after calibration on other known effective combinations. The ranking itself used network data only; its tuning drew on combinations already known to work.
Sources
Primary source: Diogo Ramos, Bruno Coutinho and Duarte Magano (University of Porto; German Aerospace Center DLR and Instituto de Telecomunicações; 21strategies), Quantum Machine Intelligence 8, article 70, published online 23 June 2026 (open access), with the authors' code and benchmark on GitHub. Also drawn on: Cheng, Kovács and Barabási on network-based prediction of drug combinations in Nature Communications (2019), Breedveld and colleagues on their adalimumab and methotrexate trial in Arthritis and Rheumatism (2006), and Ishibashi and colleagues on metformin and irbesartan in kidney tubular cells in Pharmacological Research (2012); the readiness assessment is this Monitor's own.