A*STAR Put an 18-Variable Drug-Docking Problem on Six Qubits of IBM's ibm_kingston
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · August 27, 2026.
Six qubits carried eighteen binary variables through a molecular docking calculation on an IBM quantum processor, and the selection that came back off the hardware matched the one a classical solver already had. A team in Singapore posted that result on August 20, 2026, and the number worth reading twice is the qubit count.
Practical takeaway. Nothing here reaches a clinic or a purchasing department this year. What moved is the qubit cost of a docking problem: an encoding that carries three binary variables on a single qubit, tested on real hardware against two protein-ligand complexes whose correct answer was known in advance.
What Chen, Mak and four colleagues in Singapore posted on arXiv on August 20, 2026
The preprint is titled Resource-Efficient Bio-Molecular Docking on a NISQ-era Digital Quantum Computer, by Tianqi Chen, Adrian M. Mak, Jianguo Li, Jian Feng Kong, Chandra Verma and Sebastian Maurer-Stroh. The authors work at institutes of Singapore's Agency for Science, Technology and Research, at the National University of Singapore and at Nanyang Technological University. The Quantum Insider reported the work on August 27, 2026, which is how it reached this Monitor's pool.
Molecular docking is the routine question of early drug research: given a candidate molecule and a target protein, how does the molecule seat itself in the pocket, and how well. Software answers it thousands of times a day in pharmaceutical companies and academic laboratories, because docking is cheap compared with synthesizing a compound and testing it at the bench. The difficulty is combinatorial. A flexible molecule can bend, rotate and approach the protein in an enormous number of ways, and every one of those poses is a candidate the search has to weigh.
Classical docking software handles that explosion with scoring functions and search heuristics that discard unlikely poses early. The route the Singapore team took instead is older than the current wave of quantum work: recast docking as a graph problem. Each plausible contact between the ligand and the protein becomes a vertex, weighted by the chemistry involved, whether a hydrogen bond, an electrostatic pairing or a hydrophobic patch. An edge joins two contacts that can physically coexist in one binding pose. The winning pose is then the heaviest set of vertices that are all mutually connected, which mathematicians call a maximum vertex-weighted clique. That reformulation is not the paper's contribution. What sits on top of it is.
How full-basis encoding puts eighteen binary variables on six qubits
In the standard mapping, each binary decision in the problem gets its own qubit. Eighteen possible contacts means eighteen qubits, and on machines where high-quality qubits are the scarce resource, that ratio decides what fits and what does not. The authors' answer is an encoding they call full-basis encoding. A qubit's state can be pictured as a point on the Bloch sphere, which has three orthogonal directions. They use all three to carry information, so one qubit holds as many as three separate binary variables.
The compression only works if the optimizer can still find the right answer, and the paper's supporting result is a proof rather than a benchmark: a global minimizer of the full-basis objective can always be chosen to be a pure product state. In plain terms, the optimum does not require entanglement between the qubits. Entanglement is the resource that makes quantum circuits powerful and also the one that makes them fragile, since maintaining it demands deeper circuits and more time on a noisy machine. A method whose target state needs none of it can run on shallow circuits, which is exactly what today's hardware tolerates.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The method has been run once on a real quantum processor for two textbook protein-ligand pairs with known answers, as a demonstration that the circuits execute and return the right selection; no drug program has used it, and no comparison against production docking software has been published.
What ibm_kingston returned for streptavidin-biotin and trypsin-benzamidine
The circuits were executed on ibm_kingston, a 156-qubit Heron r2 processor in IBM's fleet, with the qubit subset chosen from calibration data to hold down readout error and two-qubit gate infidelity along the layout the ansatz needs. The team ran two docking instances built from experimentally determined structures in the Protein Data Bank. The first was streptavidin with biotin, entry 1stp, which produced an 18-variable problem encoded on six qubits. The second was trypsin with benzamidine, entry 9aw2, a 14-variable problem on five qubits.
Around the encoding sits a fairly ordinary variational loop: three circuit layers, a penalty weight of four for constraint violations, a warm start of ten to fifteen Trotterized steps inspired by imaginary-time evolution, a hundred training epochs and a thousand repetitions per instance. The warm start matters more than it sounds. It begins the search in a region the classical side already considers promising, so the quantum circuit is refining a decent guess.
The reported outcome is narrow and clean. On both instances the hardware runs recovered the same vertex selections as the classical simulation and matched the known clique structure under realistic gate noise and readout error. Error mitigation was limited to standard readout mitigation, a restraint the authors describe as preserving a realistic assessment of the algorithm's robustness. Against a comparison encoding that uses only two Bloch axes, the full-basis approach reached higher success ratios, approaching certainty on the smaller trypsin instance and improving substantially on the larger streptavidin one.
The choice of those two complexes deserves a note, because it is both the strength and the ceiling of the experiment. Streptavidin and biotin form the tightest noncovalent protein-ligand pair in common laboratory use, and benzamidine sitting in the trypsin specificity pocket is the standard teaching example of a serine-protease inhibitor. The two records used here have different histories, which is worth keeping straight. Entry 1STP has been public since October 1992 and is the structure behind the classic 1989 account of biotin binding. Entry 9AW2 is a recent redetermination, released in March 2025, of a complex first characterized in the mid-1970s. In both cases the correct clique was known before the circuits ran, so the team could check the machine's answer against ground truth. A problem chosen so the answer is known is a test of the apparatus, and it says nothing yet about a target nobody has solved.
Why a problem a laptop solves in seconds is worth running on a quantum processor
The authors are unusually direct about this. They describe their hardware runs as evidence of feasibility, state that quantum advantage is not demonstrated, and point out that they deliberately stayed at graph sizes where classical verification remains possible. That candor is worth more to a reader than the result itself, because the same experiment could easily have been dressed as a breakthrough in a press release.
The value sits in the resource accounting. On a NISQ-era device the binding constraint is how many good qubits a problem consumes, and a three-to-one compression changes which problems fit on hardware that exists today. Whether that ratio survives contact with realistic docking, where graphs have hundreds of vertices and scoring functions are messier than a weighted compatibility graph, is the open question, and the authors name it as their own next step alongside better ansatz design and hardware-aware circuit compilation.
It is also useful to set this beside a very different item from the same week. On August 24 this Monitor covered a commercial screening model that ranks drug candidates in seconds with no qubits anywhere in the loop. Both are filed under quantum drug discovery in the trade press. One is a classical machine-learning product from a company with quantum in its name; the other is six qubits on a real processor solving a problem a laptop handles instantly. A clinician or an investor reading only headlines would have no way to tell them apart.
How Quentir Reads It
Three quite different things travel under the label of quantum drug discovery, and separating them is most of the work of reading this field honestly. The first is a classical model marketed under a quantum brand, where the physics is in the company name. The second is a quantum algorithm evaluated only in simulation on a classical computer, which is legitimate research and tells you nothing about whether real hardware could run it. The third is a circuit actually executed on a quantum processor, with the output checked against an answer that was independently known. This preprint is the third kind, and it says so in its own abstract.
That ordering also predicts how each kind fails. A rebranded classical model fails commercially, when a customer notices the quantum content was in the marketing. A simulation-only result fails when someone tries to compile it for real hardware and discovers the circuit depth. A hardware demonstration on a solved problem fails when the graph grows and the encoding stops paying for itself. This paper is exposed to the third failure mode and to no other, which is a comparatively good place for a piece of research to stand.
Four questions separate a serious quantum docking claim from a decorative one, and each has a short factual answer. How many qubits were used, and for how many variables. Was the circuit executed on hardware or simulated. Was the correct answer known in advance. And was the comparison run against strong classical docking software, or only against a weaker quantum encoding. The preprint answers the first three plainly and leaves the fourth for later work, which is the honest position for a result at this stage. When a vendor deck cannot answer all four, the reader has learned something anyway.
Sources
Primary source: Tianqi Chen, Adrian M. Mak, Jianguo Li, Jian Feng Kong, Chandra Verma and Sebastian Maurer-Stroh, "Resource-Efficient Bio-Molecular Docking on a NISQ-era Digital Quantum Computer," arXiv preprint 2608.19868, submitted August 20, 2026, from which the full-basis encoding, the proof that a global minimizer can be chosen as a pure product state, the maximum vertex-weighted clique formulation, the warm start inspired by imaginary-time evolution, the execution on the 156-qubit ibm_kingston Heron r2 processor with error mitigation limited to standard readout mitigation, the two test instances 1stp with 18 variables on six qubits and 9aw2 with 14 variables on five qubits, the three-layer circuits with a penalty of four and a thousand repetitions per instance, the recovery of the classical vertex selections under hardware noise, and the authors' own statement of feasibility without quantum advantage are all taken. The Quantum Insider's report of August 27, 2026 by Matt Swayne is the item through which the preprint entered this Monitor's daily pool and is the source for the plain-language framing of the docking task. The two protein-ligand complexes are identified through their Protein Data Bank entries, streptavidin with biotin at 1STP, released October 15, 1992, and trypsin with benzamidine at 9AW2, deposited in 2024 and released March 12, 2025; both release dates are read from the Protein Data Bank's own entry records rather than from the preprint, which cites the underlying crystallographic literature of 1989 and the mid-1970s instead. The comparison with a structure-free commercial screening model draws on this Monitor's post of August 24, 2026.