A quantum method for a graph-based molecular-docking problem used fewer qubits and found optimal solutions more often than a ZX-basis comparison in two tested instances, according to a new preprint. The approach, called full-basis encoding, represented up to three graph variables on each qubit. In the reported comparisons, it also recovered the same selected vertices on IBM quantum hardware as in classical simulation, matching the exact clique structures used as ground truth.
The result is a demonstration of compact encoding and hardware feasibility, not a claim that quantum computers have beaten optimized classical methods for docking. The graphs were moderate in size and could still be solved exactly with classical computation, a boundary the authors explicitly emphasize.
Packing more of the problem onto each qubit
The study recast molecular docking as a maximum vertex-weighted clique problem. In plain terms, the calculation searches for a group of mutually compatible graph vertices whose combined weight is as high as possible. The two test cases came from experimentally resolved Protein Data Bank structures: a streptavidin–biotin instance labeled 1stp and a trypsin–benzamidine instance labeled 9aw2.
The full-basis scheme used three Pauli observables on each qubit, allowing as many as three graph variables to share one qubit. The paper describes the resulting requirement as roughly one qubit for every three variables. In the matched comparison, ZX-basis encoding used approximately 50% more qubits on the tested instances.
The hardware demonstration used IBM superconducting processors, with circuits adapted through Qiskit Runtime and hardware-aware transpilation. The authors present that setup as an experimentally implementable way to test the encoding under near-term hardware constraints, rather than as a solution to every challenge in molecular docking.
Repeated tests favored the compact encoding
For the numerical evaluation, the researchers ran 1,000 randomized optimization repetitions for each approach and each docking instance. Both encodings used the same warm-start procedure, variational circuit structure, optimization protocol and training budget, making the comparison a test of the encoding under matched finite-resource conditions.
The warm start began with a few classical steps inspired by imaginary-time evolution, using matrix product states to prepare the initial conditions before the full-basis optimization. The variational ansatz used three layers and a penalty value of 4 in the main evaluation; the warm-start stage used 15 steps for 1stp and 10 for 9aw2.
The success ratio was 0.739 for full-basis encoding and 0.663 for ZX encoding on 1stp. On 9aw2, the corresponding figures were 0.993 and 0.896. Here, success means that a repetition recovered the optimal clique, so the figures describe how often the tested optimization landed on the exact graph solution under the chosen settings.
The optimization path was not simply an uninterrupted climb. The transient clique weight first fell, then rose monotonically toward the exact optimum, and the authors reported early convergence when the warm-start full-basis method was used. The pattern was observed in these two docking instances, not established as a general property of larger problems.
A feasibility result with a narrow test
The paper also reports a mathematical result about the full-basis objective: at least one global optimum of the raw objective can be represented as a pure product state, meaning the overall state can be formed from separate single-qubit states. That is a statement about the defined mathematical objective, not a measurement showing that hardware will always reach such a state.
Performance depended on the settings. Increasing circuit depth was associated with better results for 1stp, while gains saturated beyond a modest depth for 9aw2. Very small and very large penalty values were both associated with lower success, with an intermediate range performing better.
The evidence remains limited to two moderate graph instances, and exact classical verification was still possible. The study did not establish quantum advantage over optimized classical docking or graph-optimization methods, or show that the approach scales to larger, denser, more chemically realistic or flexible docking problems.
It also did not show improved binding affinity, biological activity or experimentally validated docking poses. Further work would need to test whether the qubit savings and optimization behavior persist as the graphs become larger and denser, and whether more realistic interaction models change the result.
The document is an arXiv version 1 preprint dated 20 August 2026. It reports support from National Research Foundation Singapore initiatives, with additional support from A*STAR Q. Inc and IBM Quantum Credits.
Paper data and sources
Original title: Resource-Efficient Bio-Molecular Docking on a NISQ-era Digital Quantum Computer
Authors: Tianqi Chen, Adrian M. Mak, Jianguo Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text