A Tenth of a Cent per Guess, and No Qubits in the Loop
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · August 24, 2026.
Early drug discovery spends most of its money on compounds that were never going to work. The standard defense is to guess well before anything reaches a bench, and the price of a single guess has just fallen a long way.
On August 19, 2026, SandboxAQ made a screening model called AQPotency generally available. The company's own release, issued from Palo Alto through PR Newswire, says the model 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. That is a tenth of a cent for each ranked pair.
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.
Practical takeaway. 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 itself is classical software on conventional machines.
What the model is actually asked to do
Virtual screening has one job: take a library far too large to test in glassware, and put the plausible members near the top. The older physics-based route to that ranking simulates how a molecule sits in a binding pocket, which requires knowing the shape of the pocket and buying a large amount of compute for every candidate. Machine-learned scoring is cheaper and has a long history of flattering itself on benchmarks that resemble its training data more than they resemble a new target.
SandboxAQ calls AQPotency a Large Quantitative Model, its house term for a model trained on physical and chemical quantities rather than on text. The release pairs the launch with general availability of a second such model for catalyst discovery, and makes AQPotency reachable both through the company's own site and through an assistant integration, with a cloud marketplace listing promised afterward. The distribution detail matters more than it looks: a screening model a chemist can call from an ordinary working session is a different object from one that needs a computational chemistry group to operate.
Andrea Bortolato, the company's vice president of drug discovery, told the quantum industry outlet Quantum Zeitgeist in its account of the launch that the tool gives customers a way to prioritize compounds without needing a three dimensional crystal structure of the target. Dario Alessi, who directs the Medical Research Council Protein Phosphorylation Unit at the University of Dundee, says in the same record that SandboxAQ models let his Parkinson's team explore a much larger biochemical space in a short timeframe and improve both activity and selectivity.
The number that makes it interesting
A tenth of a cent per comparison changes what a screening campaign is. 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.
The caveat is that 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 the named academic collaborations. Those are the company's own account of its own tool. No independent evaluation, no prospective blinded benchmark and no head to head against an established free-energy method appears in the material released with the launch.
Quantum pillar: not applicable. Technology readiness: TRL 9 of 9. The company reports the software as generally available and in use across eight customer programs, which is the top rung for a working research tool, and that rung rests on the vendor's own account of deployment rather than on any independent evaluation.
The Parkinson's result, read from its own preprint
The launch material also carries a quote from Gary Miller at Columbia University's Mailman School of Public Health about selective binders for SV2C, a synaptic vesicle protein enriched in dopaminergic neurons of the basal ganglia and implicated in Parkinson's disease. That campaign has a record of its own. On the same day as the launch, a preprint titled Discovery of Selective Small-Molecule Ligands of SV2C by AI-Enhanced Virtual Screening and Experimental Validation appeared on bioRxiv with Miller as its final author. Its corresponding author, Alexander Brueckner, gives SandboxAQ in Palo Alto as his institution, and Bortolato is among the co-authors, so this is a company reporting on its own methods alongside academic collaborators.
Read the methods and the picture sharpens in an instructive way. Because no full-length high-resolution SV2C structure existed, the team built a homology model from SV2A cryo-electron microscopy templates, explored its conformations with molecular dynamics, and applied a convolutional neural network scoring function, retrospectively validated against a curated 39-ligand SV2A benchmark at a correlation of 0.72, in a funnel that narrowed 5.96 million in-stock commercial compounds to 3.19 million with central nervous system relevance before docking and rescoring. Of 94 prioritized candidates, 71 were profiled experimentally and 22 were active, a 31 percent hit rate. Two leads, numbered 36 and 56 in the paper, bind SV2C with inhibition constants of 24.6 and 3.25 micromolar, both with more than tenfold selectivity over the related isoform SV2A.
That is a real result on a target that had no selective small-molecule probes at all. It is also not an AQPotency run, and it did not proceed without a structure: the pipeline built one when none existed, then docked into it. The launch announcement and the preprint are two different kinds of evidence about the same company, and only the second one carries a methods section a reader can check.
Where the quantum is, and where it is not
Set all of this beside what a quantum computer can currently do with a molecule. On August 20, one day after the AQPotency launch, Yingrong Chen, Nathan Baker, Rushi Gong, Conrad Johnston, Brad Lackey, Hongbin Liu, Sasha Schmidt, Yuan Su, David Williams-Young and Yinuo Yang posted a preprint on sparse quantum state preparation for molecular wavefunctions. Their subject is the step before any chemistry happens: loading a description of the molecule onto the qubits at all.
The work is careful and the improvement is real. By exploiting affine relationships among electronic configurations over the two element finite field, the authors compress the register using only Clifford gates and no helper qubits, then trade a few Toffoli gates for further compression when it pays. Across their molecular benchmarks the method needs the fewest ancillary qubits of the approaches they evaluated while holding the expensive gate count comparable. Ancillary qubits are the scratch space a circuit borrows to do its work, and on hardware where every qubit is scarce, needing fewer of them is worth a great deal.
Notice what the two records describe. One is a shipped product priced in fractions of a cent, with a same-day preprint showing a classical pipeline delivering micromolar hits on a hard membrane target. The other improves the cost of the first step of a calculation that no existing machine can yet carry to a useful answer. Both belong in a quantum medicine pool, and they sit at opposite ends of the readiness ladder. This Monitor read the same distance from the other side yesterday, in a paper that cut a quantum error rate on a test circuit containing no molecule.
The confidence interval is the clinical part
One feature of the launch deserves more attention than the price. AQPotency reports, for every prediction, how confident it is and whether the target falls inside the range where the model performs reliably. That second quantity has a formal name in cheminformatics, the applicability domain, and it is the thing that separates a usable model from a plausible one.
A scoring model trained on the chemistry that exists will be confident about chemistry that resembles it and quietly wrong about chemistry that does not. A single number with no uncertainty attached gives a medicinal chemist no way to tell those two situations apart, which is how screening tools acquire their reputation for wasting a year. A model that declines to be sure, and says so in a machine readable way, can be audited. It also lets a program record why a compound was chosen, which is the kind of trace that matters later when a regulator or a partner asks how a candidate was selected.
For a hospital, AQPotency is purchasable today and still changes nothing on a formulary. It sits three or four handoffs upstream: a ranked compound becomes a synthesized compound, which becomes a preclinical candidate, which may become a trial. The value of watching it here is that the upstream tooling shapes which diseases get programs at all, and cheap scoring pushes on exactly the targets that were previously abandoned for want of a map.
How Quentir Reads It
The lesson of this pairing is about labels. A quantum medicine pool will surface a quantum-branded company shipping a classical tool and an academic group doing genuine quantum work on a problem that remains far from a molecule of clinical interest. Reading either one by the company's sector rather than by what the technology does produces a wrong answer, and it produces a wrong answer in opposite directions.
The readiness ladder is the instrument that keeps them apart. AQPotency sits at the top of it because the company reports the software as finished, distributed and in use, and that rung says nothing whatsoever about whether the medicines it helps choose will work. A quantum chemistry preprint sits near the bottom of it and may still be the more scientifically important of the two in a decade. A hospital buyer who is told that a supplier uses quantum methods for discovery has one useful question available: which part of the pipeline, on which hardware, and what would the classical baseline have cost.
The Parkinson's work is the part worth holding on to. SV2C is a hard membrane target in a disease with no disease-modifying therapy, and micromolar selective binders where none existed are a genuine contribution to a field that has been stuck. That contribution now rests on a posted preprint with a full methods section, written by a team whose corresponding author works for the company whose software it evaluates, which is a reasonable place to start and not a place to stop. Peer review, an independent replication, or a compound from one of those eight customer programs entering a formal preclinical package would each move it further. Until then, the accurate description of August 19, 2026 is that virtual screening got materially cheaper and rather easier to reach, on conventional computers, from a company most people file under quantum.
Sources
Primary source: SandboxAQ, "SandboxAQ Launches AQPotency, Bringing Ultrafast Virtual Screening to Drug Discovery Without a Solved Protein Structure," company release distributed by PR Newswire, Palo Alto, August 19, 2026; pricing, speed, hardware, applicability-domain reporting and the count of customer programs read from that release. The SV2C campaign is read from its own preprint by Alexander Brueckner and colleagues, corresponding author at SandboxAQ, with Gary Miller as final author, posted to bioRxiv on August 19, 2026; methods, funnel sizes, hit rate and binding constants come from its abstract. Bortolato and Alessi quotes as carried in Quantum Zeitgeist's August 20, 2026 account of the launch. Quantum state preparation comparison: Yingrong Chen and colleagues, arXiv preprint, August 20, 2026.