A*STAR Put an 18-Variable Drug-Docking Problem on Six Qubits of IBM's ibm_kingston
Six qubits for an eighteen-variable docking problem
On August 20, 2026 six researchers at institutes of Singapore's Agency for Science, Technology and Research, the National University of Singapore and Nanyang Technological University posted a preprint describing a hybrid quantum-classical method for molecular docking, and executed its circuits on IBM's ibm_kingston processor. Molecular docking is the everyday 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. The team recast that question as a graph problem, where each plausible contact between ligand and protein is a weighted vertex and the winning binding pose is the heaviest set of contacts that can all coexist.
An encoding that carries three variables on one qubit
The contribution is an encoding the authors call full-basis encoding, which uses all three orthogonal directions of a qubit's Bloch sphere to carry information instead of one. That put an 18-variable problem on six qubits for the streptavidin-biotin complex 1STP, and a 14-variable problem on five qubits for trypsin with benzamidine, entry 9AW2. The paper also proves that a global minimizer of its objective can always be chosen to be a pure product state, so the optimum requires no entanglement between qubits and the circuits can stay shallow enough for current hardware. What changed for a reader tracking this field is the qubit cost of a docking problem, since the number of good qubits a problem consumes is the binding constraint on machines available today.
What the hardware runs did and did not settle
On both instances the runs on ibm_kingston recovered the same vertex selections as the classical simulation and matched the known clique structure under realistic gate noise and readout error. The authors state plainly that this is evidence of feasibility and that quantum advantage is not demonstrated, since both test problems were deliberately kept small enough for classical verification. Streptavidin with biotin and benzamidine in the trypsin pocket are textbook complexes, though their records differ in age: entry 1STP has been public since 1992, while entry 9AW2 is a 2025 redetermination of a complex first characterized in the mid-1970s. Either way the correct answer was known before the circuits ran, which is what made the check against ground truth possible and also what limits the claim. The open question the authors name themselves is whether the compression ratio survives realistic docking graphs with hundreds of vertices, alongside better ansatz design and hardware-aware circuit compilation.
A Tenth of a Cent per Guess, and No Qubits in the Loop
A tenth of a cent per ranked pair
On August 19, 2026, SandboxAQ made a virtual screening model called AQPotency generally available. The company's release says it scores how strongly a candidate molecule is likely to act on a disease target, ranks molecule and target pairs in seconds, runs on ordinary computing hardware, and costs as little as one dollar per thousand comparisons. At the advertised minimum, a million pairwise rankings come to about a thousand dollars. Screening at that price stops being a budgeted event a team plans around and becomes something a chemist can run while thinking.
Screening without a solved protein structure
The claim that will interest a hospital pharmacologist sits one line further down. AQPotency is described as working without a solved protein structure, which makes it structure-free potency prediction: the model does not need a crystallographic or cryo-electron map of the target before it will score anything. A great many disease targets have no such map, and programs aimed at them have historically stopped at that wall. Cheap ranking only helps if the ranking is right, and the release publishes no accuracy figures. What it offers instead is eight customer programs with what the company calls experimentally validated impact, plus named academic collaborations. No independent evaluation, no prospective blinded benchmark and no head to head against an established free-energy method appears in the launch material.
The Parkinson's campaign has its own preprint
The launch quotes Gary Miller of Columbia University's Mailman School of Public Health on selective binders for SV2C, a synaptic vesicle protein implicated in Parkinson's disease. That campaign was posted to bioRxiv the same day, with a SandboxAQ corresponding author and Miller as final author. Its methods matter: because no full-length SV2C structure existed, the team built a homology model from SV2A cryo-electron microscopy templates, ran molecular dynamics, and applied a convolutional neural network scoring function inside a funnel that narrowed 5.96 million commercial compounds to 3.19 million before docking. Of 94 prioritized candidates, 71 were profiled and 22 were active. It is a genuine result on a target with no selective probes, and it is neither an AQPotency run nor a structure-free one.
The pillar field reads not applicable
Nothing in this announcement runs on a quantum computer. The most operationally ready development in today's quantum medicine pool comes from a company the quantum trade press covers closely, and the tool is classical software on conventional machines. One day later, a separate group posted a preprint on loading a molecule's wavefunction onto qubits at all, which is the step before any quantum chemistry happens. Both belong in the pool, and they sit at opposite ends of the readiness ladder.
When the Error Check Happens Decides What It Buys
A 54 percent error cut, and what it was measured on
Simulating a molecule on a quantum computer means chopping its time evolution into many short slices, a technique called Trotterization. Accuracy improves as the slices get finer, and finer slices make a deeper circuit, and every additional gate in that circuit is another chance for a physical fault. Depth is what the chemistry asks for and depth is what today's hardware punishes, which is the bottleneck the field calls the deep Trotter dilemma.
A preprint from IonQ, qBraid and NVIDIA, posted in May 2026 by James Brown, Jason Iaconis, Yuri Alexeev and colleagues, reports that a combination of the Generalized Superfast Encoding, Clifford Noise Reduction and Shor-style stabilizer verification lowered the logical error rate by up to 54 percent on a trapped-ion machine. The encoding itself carries no such requirement, and Clifford Noise Reduction can be run with its stabilizer readout deferred; what the paper establishes is that the advantage over the unprotected baseline depended on mid-circuit measurement, the ability to read some qubits partway through a run, learn from them and continue with the rest.
What the number does and does not cover
The unprotected baseline is a six-qubit encoded Clifford Trotter step; the protected implementation that produced the 54 percent figure is wider, at 26 qubits and 580 gates, the extra width being the verification machinery itself. Both ran on a Barium development system similar to IonQ's forthcoming Tempo line. Clifford circuits are the well-behaved subset of quantum operations an ordinary laptop can simulate exactly, which is what makes them useful test articles: the correct answer is known in advance, so any deviation is measurable error. No molecule, binding energy or reaction barrier appears anywhere in the result, while IonQ's own account presents the same figure alongside the prospect of lower research costs and shorter time to market.
The control arm is the transferable finding
Keep the verification structure but move the stabilizer readout to the end of the circuit and it still beats having no error detection at all. What it stops doing is beating the unprotected baseline by a statistically significant margin for a single stabilizer round. Only the mid-circuit version clears that bar, which points at the timing of fault detection as the operative ingredient. For scale, published resource estimates for computing spin gaps in models of the cytochrome P450 catalytic cycle, a methodology benchmark and not a dosing or interaction predictor, run near 1,434 logical qubits and roughly 4.6 million physical qubits over 73 hours, on stated and assumption-dependent compilation figures. This Monitor records the work at TRL 3 on the simulation pillar.
A Paid Quantum Drug Project Meets a Four-Stage Validation Ladder
The next calculation became funded work
China Securities Journal reported on August 14 that Shenzhen Jingtong Life Science tested Boson Quantum's approach on a drug-target structure problem in June and moved to a paid quantum drug-discovery engagement in July. The client named two difficult tasks: finding the precise binding conformation of a drug-target complex and calculating its energy after binding. It also reported an advantage in conformation search. The work now covers new molecules associated with healthy aging. The article says drug-related intellectual property would be jointly owned if the project succeeds, while software and algorithm rights would remain with Boson. The account is commercially meaningful because a buyer has funded a defined computational problem, yet the public article provides no molecule list, hardware configuration, classical baseline, runtime, success metric, or laboratory confirmation.
The same report sets a demanding standard
GuoDun Quantum executive Wang Zhehui supplied a four-stage validation ladder in the same article. It begins with benchmark advantage and initial industry validation on real quantum hardware. The decisive stages are superiority to classical algorithms under equal conditions and reproducible business outcomes. The engagement is commercially concrete, but the disclosed record cannot yet assign it a validation rung because real-hardware use and a controlled comparison remain undisclosed. Funded work may support learning, access, feasibility, or method development. It establishes customer demand without settling the scientific comparison.
The medical handoff remains ahead
Quantum computation would sit near the beginning of drug development. It may help decide which molecules deserve synthesis and testing, though it cannot establish safety, dosage, biological effect, or clinical benefit. Quentir reads the account as a useful commercial milestone with a built-in limit. The reported planned IP allocation recognizes a reusable computational method on one side and, if the project succeeds, a medicine that must survive laboratory, preclinical, and clinical work on the other. Paid work has started before a reproducible pharmaceutical result. The next public milestone that changes the judgment would connect an equal-condition computational comparison to a wet-lab result that another team can understand and repeat.
A Photosensitizer Has Two Jobs Before Quantum Computing Helps
Two molecular jobs define the target
A light-activated cancer drug must absorb light in a useful therapeutic window and turn that energy into chemistry that damages a tumor. Xanadu and the University of Alberta announced on August 13 that they will develop quantum algorithms for photosensitizer design. The research partnership follows a December 2025 preprint that specifies both desired outputs: cumulative absorption and intersystem crossing rates linked to reactive-oxygen production. The paper applies proposed fault-tolerant algorithms to BODIPY derivatives, including substitutions that are difficult for standard computational chemistry. This is a sharply defined pharmaceutical calculation, although the announcement reports no new molecular result, hardware run, candidate compound, or development timeline.
The resource estimate belongs to a future machine
The preprint studies active spaces of 11 to 45 spatial orbitals and estimates a need for roughly 180 to 350 logical qubits. Logical qubits are encoded across physical qubits and require error-correction overhead, so that estimate cannot be compared directly with headline physical-qubit counts. Xanadu therefore places the work in fault-tolerant quantum computing, while Professor Alex Brown contributes expertise in photodynamic therapy chemistry and the excited-state processes that classical methods struggle to capture. The partnership joins an algorithm team and a chemistry group around a named failure point in simulation. Its technology readiness remains early because no calculation has yet been reported on suitable hardware or checked against laboratory measurements.
Clinical usefulness still runs through the laboratory
Photodynamic therapy is an established local treatment for selected cancers and precancers, but better photosensitizer calculations must still survive pharmacology, toxicology, formulation, delivery, and trials. A future quantum result would need comparison with measured spectra, photochemical rates, and strong classical baselines. Quentir reads the partnership as a disciplined research bet: the biomedical quantities are prespecified, the molecular class is named, and the resource assumptions are public. Progress will become legible when the workflow reproduces a known molecule's behavior and then ranks unfamiliar candidates well enough to save synthesis cycles.
A Korean Pharma Webinar Names the Quantum Workload
The announcement names the computing path
KPBMA's first AI drug-discovery webinar, focused on quantum computing, is scheduled for August 13. Tech42 reports that SDT plans to present a hybrid quantum computing stack built around its QuREKA service. The announced environment links CPUs, GPUs, simulators, and quantum processors through a CUDA-Q development layer. Participants are also expected from QuantWare, Quantum Intelligence, and the Korea Research Institute of Bioscience and Biotechnology. The public account names molecular simulation, candidate search, and ADMET prediction as possible uses. Those are concrete pharmaceutical tasks, but the source remains an event preview. It provides no methods paper, dataset, benchmark table, candidate molecule, or measured advantage.
Portability can expose what the QPU contributes
The useful feature may be the ability to run one drug-discovery workload across different backends. A common development environment can keep the surrounding classical code visible while a team compares a simulator, GPU, and available QPU. That does not establish better performance. It can make the comparison easier to inspect. Runtime, accuracy, resource use, variance, and sensitivity to hardware noise all become part of the result. Pharmaceutical researchers can then ask whether the quantum step improves a defined endpoint or only changes the route used to compute it.
The pharmaceutical result is still open
ADMET covers how a drug is absorbed and distributed, how it is metabolized and excreted, and whether it is toxic. Each property depends on different data and validation methods. SDT's preview does not yet name an endpoint, baseline, or test set. Quentir therefore records two separate steps: SDT has described a multi-backend architecture, and KPBMA has created a pharmaceutical forum for it. Readiness for the announced applications remains unranked until a reported experiment shows what ran on the QPU, what stayed classical, and how the output compared with a strong existing method.
Inside the 156-Qubit Enzyme Calculation
QC Ware reports a molecular calculation on quantum hardware
QC Ware says it calculated the electrostatic interaction energy of nitric oxide reductase by combining GPU-accelerated molecular modeling, classical chemistry methods, and quantum measurements on IBM's 156-qubit Heron processor. The enzyme is chemically demanding because its active region contains metal. The public announcement makes the electrostatic interaction energy calculation concrete, but it does not disclose the molecular partition, circuit design, measurement count, error mitigation, reference value, or final numerical error. The release therefore supports a precise statement: QC Ware reports that a medically relevant class of molecular calculation reached named quantum hardware. It does not yet show that the quantum step improved the result.
The architecture has a clear division of labor
The hybrid chemistry workflow combines Promethium, GPU-accelerated molecular modeling, classical chemistry methods, and quantum measurements. The release does not disclose how work was partitioned among them. IBM's published description of Heron and System Two provides useful architectural context: its quantum processors operate with classical runtime servers and methods that divide larger calculations. The QC Ware release is the source for the later 156-qubit hardware claim. It also says the demonstration is not currently an integrated Promethium product capability. That sentence prevents a hardware claim from being mistaken for a production service.
The missing benchmark defines the next milestone
The announcement reports no quantum advantage and offers no comparison against a strong classical workflow for the same chemical task. Qubit count cannot supply that missing result. A buyer would need comparative accuracy, resources, runtime, repeatability, and a decision consequence for chemists. The public record therefore places the work at TRL 3: a vendor-reported hardware proof of concept for one molecular property. Its product boundary is commercially informative because it separates an experimental module from the platform available today. The next persuasive record would show what the quantum measurements add at a fixed cost or error, and whether that contribution changes a research decision.
A Drug-Design Race With Two Different Finish Lines
Two systems pursue different molecular virtues
A 2026 Artificial Intelligence Chemistry paper compares a quantum-aided molecular generator with BInD, a diffusion model for structure-based drug design. QuADD treats molecular design as constraint-driven optimization: it searches for candidates that fit a defined binding pocket while satisfying several selected properties. BInD uses reverse diffusion to explore possible molecular structures and their interactions with the target. Both generated novel candidates for a thrombin binding site. The reported split matters. QuADD more consistently met the study's criteria for predicted binding affinity, drug-likeness, and preservation of key protein-ligand interactions. BInD produced greater structural diversity. That result describes two search philosophies as much as two software systems.
The clock measures a complete workflow
A company-supplied account reports that QuADD generated 3,000 molecules in roughly 30 minutes, while BInD required about 40 hours on a node with one NVIDIA GPU. The researchers selected the top 100 candidates from each set using predicted binding affinity. The timing is useful as a workflow observation. A broader quantum-advantage claim would need matched objectives, declared compute resources, complete timing boundaries, repeated runs, and a strong classical optimizer aimed at the same constraint set. The comparison article was written by researchers affiliated with Polaris Quantum Biotech, the company behind QuADD, so independent reproduction would add weight.
The medical threshold remains physical
The candidates in this study were ranked by computational proxies. The public record reports no synthesis campaign, measured binding assay, cellular result, toxicology study, or clinical outcome. Those later tests decide whether a promising structure can become a useful lead. This analysis follows the line from objective function to patient relevance: broad exploration can reveal unfamiliar scaffolds, while constrained search can reduce the number of weak or impractical candidates sent to the laboratory. A productive discovery pipeline may combine both. The immediate result is narrower and still worthwhile. In one thrombin task, the quantum-aided system prioritized the selected properties more consistently, while the diffusion model searched more widely. The next consequential step is a transparent, independently repeated benchmark joined to physical molecules and measured biology.
Why a 12,635-Atom Protein Simulation Still Needs Supercomputers
A protein-scale calculation divided across machines
A May 2026 arXiv preprint joins quantum processors with the Fugaku and Miyabi-G supercomputers to model trypsin and T4 lysozyme with ligands and surrounding water. The larger system is a 12,635-atom protein-ligand complex. That total describes the biological scene, not one enormous quantum circuit. The researchers used embedding to divide each molecule into electronic fragments, assigning selected difficult fragments to quantum hardware while classical systems prepared and assembled the wider calculation. This heterogeneous quantum-classical workflow is the central result: current quantum processors can participate in a large biomolecular simulation when their task is bounded carefully and supported by substantial classical computation.
What the resource record establishes
The team used two IBM Heron r2 processors and up to 94 qubits in individual calculations. It ran 9,200 circuits for more than 100 hours, collected 1.3 billion measurement outcomes, and processed the sampled data on two supercomputers. The paper reports a system more than 40 times larger than an earlier 303-atom demonstration and up to 210 times better accuracy than a previous quantum-centric approach in one workflow step. Selected fragment energies matched a respected coupled-cluster classical reference. IBM also states that the complete method still trails leading classical approaches.
Why drug discovery still has another threshold
The study establishes enabling infrastructure for biomolecular simulation. It does not report a prospective compound-selection campaign, blinded prediction, wet-lab confirmation, or a medicine advanced because of the calculation. Drug discovery needs reliable differences between candidate molecules, not an atom total alone. A later benchmark could show whether this method ranks a difficult ligand series more accurately or predicts an experimental energy difference that practical classical approximations miss. That is the bridge from a large computing experiment to a laboratory decision. Quentir reads the current work as a serious architecture milestone whose most useful feature is its transparent account of quantum execution, classical support, comparison limits, and remaining medical distance.
A quantum sampler leaves the whiteboard
Drug discovery often begins with a search through more molecular possibilities than any laboratory could ever test one by one. A quantum version of a familiar sampling routine has now run on Quantinuum's H2 and Helios systems, producing accurate results on physical qubits in a tightly bounded experiment. The practical lesson is demanding rather than triumphant: useful sampling methods must survive hardware noise before any theoretical speedup can help real molecular research.
The method at the center of this work is Markov Chain Monte Carlo, a workhorse for drawing samples from complicated probability distributions. In chemistry, those distributions can describe the many configurations a molecule may adopt, and the questions that matter often reduce to an average taken over that vast space. The appeal of a quantum approach is specific and bounded: quantum amplitude estimation offers a quadratic reduction in the resources needed to estimate certain averages, provided the machine can first prepare the right probability distribution. That proviso has always been the awkward part, and it is exactly what this experiment set out to test on real hardware.
In a March 2026 preprint, Baptiste Claudon, Sergi Ramos-Calderer and Jean-Philip Piquemal encoded two-state Markov chains, prepared their stationary distributions, and ran the algorithm on Quantinuum's H2 and Helios computers, within a collaboration between the Centre for Quantum Technologies in Singapore and Qubit Pharmaceuticals. They keep the claim modest: the experiment uses the simplest non-trivial chains and tests the building blocks of the method, not a pharmaceutical molecule. That restraint is what makes it useful. It isolates the sampling machinery, shows which pieces can already survive a real device, and hands researchers something concrete to improve next: state preparation, circuit depth, error behavior, and the handoff between quantum sampling and classical analysis. For medicine, the humane stake arrives much later, after years of chemistry, toxicology and clinical work, so the near-term value is scientific discipline rather than a faster cure. A hardware run with clearly stated limits is more useful to a decision-maker than a grand promise, because it lets the field see exactly how far the computation has traveled and how far it still has to go, one reproducible step at a time.