Researchers Reviewed 53 Quantum Computing Studies in Bioinformatics. Two-Thirds Never Used a Quantum Computer.
Table of Contents
September 24, 2026 – Five researchers from two Italian universities published a systematic review of 53 peer-reviewed experimental studies that applied quantum computing to bioinformatics and biomedical data. The review, in Frontiers in High Performance Computing, found that 36 of the 53 studies never ran on quantum hardware.
Forty of the studies tested a quantum or quantum-inspired method against at least one classical method, and the authors found that all 40 reported results comparable to or better than the classical baseline. They wrote that none of the studies they examined explicitly demonstrated a quantum advantage over classical approaches.
The authors are Manuel Arcieri and Paolo Zuliani of the Department of Computer Science at Sapienza University of Rome, and Giovanni Carotenuto, Meryam Carrus and Tiziana Castrignanò of the Department of Ecological and Biological Sciences at the University of Tuscia in Viterbo. Zuliani and Castrignanò are the corresponding authors.
The team searched the Scopus database, which returned 1,550 records, and excluded preprints, papers published before 2014, purely theoretical proposals and papers with no Scopus citations unless they appeared after January 2025. In equivalent searches of PubMed, Web of Science, IEEE Xplore and the ACM Digital Library, the authors found no further qualifying studies.
The 53 studies cover genome assembly and sequence alignment, protein structure and drug discovery, phylogenetics, and quantum machine learning for biomedical classification such as cancer genomics, disease prediction and medical imaging.
Twenty-five studies, or 47%, were classical simulations of quantum circuits, and another 11 used quantum-inspired algorithms, which borrow ideas from quantum computing but run entirely on conventional computers. Eleven ran on quantum annealers and six on gate-based quantum processors.
All 11 annealing studies used D-Wave machines from three generations, the D-Wave 2X, the D-Wave 2000Q and the first-generation Advantage. The six gate-based studies ran on processors from IBM, IonQ, Rigetti and QuantumCTek.
The authors classified 41 studies, or 77%, as tested only on toy or substantially reduced datasets, and 12 as using data close to real-life scale. Fifteen of the 17 studies that ran on quantum hardware used reduced datasets.
Eighteen studies published their source code and 37 gave access to their experimental data. Sixteen, or 30%, provided both, which the authors called the minimum requirement for full computational reproducibility, and 14 provided neither.
The authors listed obstacles specific to quantum experiments, among them costly and restricted access to quantum processors, hardware calibration that drifts over hours to days, opaque vendor software stacks and results that depend heavily on settings such as annealing schedules. They noted that the D-Wave 2X and 2000Q systems used in several of the reviewed studies are no longer commercially available.
They recommended standardized benchmark datasets and evaluation protocols, outlets for negative results, public deposit of code, data and hardware settings, measurement of how methods scale with problem size, domain-specific software libraries, and labels that separate hardware experiments from circuit simulations and quantum-inspired classical heuristics.
The authors placed the paper alongside a 2026 systematic review by Khalaj, Rayan and Jin in Briefings in Bioinformatics, and said they had focused specifically on experiments run on real devices, on classical baselines and on reproducibility.
My Analysis
In my quantum machine learning scorecard in May I proposed five questions for any quantum-AI headline. The second was whether the result came from real hardware or a simulator. Arcieri and colleagues have now put that question to the 53 peer-reviewed experimental studies that met their selection rules. For 36 of them, the answer is a simulator or a classical algorithm.
Within this corpus, 27 of the 40 studies that reported matching or beating a classical method never ran on quantum hardware, and 29 of the 40 used reduced datasets. That is 67.5% and 72.5%. I’d use those proportions as a starting assumption, not a prediction. When a paper or a pitch says quantum computing matched or beat classical methods on a biology problem, I would assume a simulation or a toy dataset until the authors state otherwise.
Simulations and Quantum-Inspired Methods Outnumber Hardware Runs Two to One
A simulation of a quantum circuit is a classical computation, and a quantum-inspired algorithm is a classical algorithm that borrows representations or techniques from quantum computing. Both have uses. Algorithm designers use simulators to check circuits and estimate resources before paying for processor time. Some quantum-inspired heuristics compete well with conventional methods. Neither shows how a physical processor performs once noise, sampling and run time are counted. The authors put the practical limit of circuit simulation on conventional clusters at roughly 20 to 40 qubits. Within that limit, researchers can simulate only small circuits, however large the dataset behind them.
In my Quantum AI taxonomy, quantum-inspired software on classical hardware is one of seven different things people mean when they say “quantum AI”. I argued there that the label is fair when the author says what hardware ran the code. Arcieri and colleagues ask researchers for the same disclosure in their fifth recommendation. Readers who rely on titles and press coverage depend on that disclosure to tell a hardware experiment from a simulation.
Reported Losses Inside the 40 Parity-or-Better Studies
Arcieri and colleagues say they assessed the quality of each classical comparison, but their summary tables record only whether a baseline was present and include no rating of how strong or well tuned each baseline was. Baseline strength was the third question on my list. The authors also state that none of the 40 studies demonstrates a quantum advantage. Those studies measured the quality of the answers their methods produced, and 29 of the 40 did so on reduced datasets.
The most striking claim I found in the corpus comes from Yuqi Zhang and colleagues, who ran a variational quantum eigensolver on the IBM Eagle processor installed at the Cleveland Clinic, a machine with 127 physical qubits. They reported lower structural error and better docking scores than AlphaFold3, the structure-prediction model from Google DeepMind and Isomorphic Labs. Their benchmark was short peptide fragments from ligand-binding sites, which they describe in their abstract as a setting where conventional methods often fail. Arcieri and colleagues class the dataset as toy or reduced. Zhang and colleagues credit AlphaFold2 and AlphaFold3 with their success on full-length proteins.
Across the 23 benchmark fragments, the quantum pipeline had the lower structural error on 18 and AlphaFold3 on the other five, while the quantum predictions scored better on docking for 21. Docking scores are a computational proxy for binding. Fragments that conventional methods often get wrong are also a long way from whole proteins.
Zhang and colleagues report those five losses openly. So do Richard Li and colleagues, whose 2021 cancer-classification study I return to below: with the full training data, standard machine learning outperformed quantum annealing in every two-class comparison. As I read the tables, Arcieri and colleagues classified each paper by its overall conclusion and counted the Zhang and Li studies among the 40. That way of counting is a separate issue from publication bias, which the authors raise in their meta-analysis and then, in their discussion, say they found no evidence of.
The One Study With Quantum Hardware, Realistic Data, a Baseline and Open Code
I filtered the review’s four summary tables by the criteria the authors applied and relied on their classification of each study. Only one of the 53 studies combines quantum hardware, data rated close to real-life scale, a classical baseline, and published code and data. It is the 2021 paper in Patterns by Li and colleagues at the University of Southern California and Genuity Science, who classified cancers from The Cancer Genome Atlas on D-Wave 2X and 2000Q annealers. Li and colleagues described the method in their title as inspired by quantum processors.
For machine-learning studies, Arcieri and colleagues count a full clinical cohort as real-life scale even when the researchers compress it with classical methods before the quantum step. To fit the annealers, Li’s team compressed tens of thousands of molecular features into 44 principal components for the two-class comparisons and 13 for the six-cancer comparison, the most the hardware could accommodate.
The paper’s Ising-type methods, which pose classification as an energy-minimization problem that either an annealer or a classical solver can handle, did better than standard machine learning when the training data was cut to small fractions. In several comparisons, simulated annealing on a conventional computer matched or beat the D-Wave machines, and so did a control that generated 1,000 random candidate solutions and combined the 20 with the lowest energy.
Li and colleagues wrote that the benefit of quantum annealing “cannot be attributed solely to inherent quantum behavior”, and they credited the Ising formulation of the problem, which simulated annealing also solves on an ordinary computer. The test I would put to any quantum pipeline is the one Li and colleagues ran on their own work: replace the quantum step with a strong classical solver, keep everything else the same, and compare.
The other hardware study with data rated close to real-life scale is a 2022 docking study of a lung surfactant protein and the SARS-CoV-2 spike, whose authors ran the quantum step on a Rigetti processor to solve MaxCut problems on subgraphs of three atoms each. Arcieri and colleagues rate it as sharing neither code nor data; its authors offer raw data on request.
Reproducing the Hardware Studies
Six of the 17 hardware studies ran on the D-Wave 2X or 2000Q. Arcieri and colleagues list both machines as no longer commercially available and say that literal reproduction of those experiments is therefore impossible.
Where the hardware still exists, cost is the first obstacle the authors list. In the arXiv version of their QDockBank dataset paper, which they presented at the SC ’25 supercomputing conference, Zhang and colleagues reported that predicting 55 protein fragments took tens of hours on superconducting quantum processors at a total computational cost above one million US dollars. Published code, data and hardware settings are the practical starting point for checking work like that. Two-thirds of the 53 studies did not publish their code.
Quantum optimization researchers now have a shared benchmark library, which Koch and colleagues published in Nature Computational Science this year. Arcieri and colleagues say they know of no equivalent for bioinformatics, and building one is their first recommendation.
Fault-Tolerant Machines and the Life-Sciences Timeline
Arcieri and colleagues list four conditions for practical quantum advantage in bioinformatics: enough high-fidelity logical qubits to encode real problem instances, logical error rates low enough for the circuit depths involved, connectivity overheads that leave the theoretical gain intact, and end-to-end runtimes that beat optimized classical and supercomputer baselines at the same accuracy. They also note that fault tolerance is widely believed necessary for practical advantage in applications such as bioinformatics. Every hardware study in their corpus ran without quantum error correction, on machines from what John Preskill named the noisy intermediate-scale quantum, or NISQ, era in 2018.
In my assessment of each industry for 2033, I rated pharmaceuticals and biotechnology among the sectors with the highest competitive impact because of molecular simulation on fault-tolerant machines. Published estimates for one widely studied model of the cytochrome P450 active site, from the enzyme family that metabolizes most marketed drugs, run from roughly 1,000 to 4,900 logical qubits depending on the method.
In its June 2025 roadmap, IBM set 2029 as the target for Starling, a system planned to run 100 million gates on about 200 logical qubits. Simulating molecules at the P450 scale is a different workload from the optimization, search and classification experiments that make up most of this corpus.
Arcieri and colleagues call the field’s current output “credible prototypes”, and I agree. Researchers will carry the encodings, software and benchmarks they build on today’s machines to fault-tolerant hardware. Other groups can test and reuse that work most easily where code and data are public, as they were for 16 of these 53 studies.
