Kipu Quantum Ran a Molecular Toxicity Classifier on QuEra's 256-Qubit Aquila: What 200 Neutral Atoms Added for 122 Drug-Candidate Molecules
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · October 17, 2026.
The German quantum software company Kipu Quantum has reported a molecular toxicity experiment run on QuEra Computing's Aquila, a neutral-atom quantum processor with up to 256 atoms. Each of 122 molecules from a public toxicity dataset was encoded into the controlled quantum-state evolution of 200 atoms, and the measured patterns were turned into about a thousand new numerical features per molecule for an ordinary machine learning classifier.
On those 122 molecules, drawn from a dataset of 171 laboratory-tested compounds, the quantum features raised the balanced accuracy of the toxic or non-toxic call from about 0.55 to about 0.70, according to the authors. For anyone who screens drug candidates for safety, the result is an early test of whether a quantum device can make a classical prediction model better at spotting the harmful minority of compounds.
The work is described in a technical blog post on the company's site, Reading Toxicity of Molecules with Neutral-Atom Quantum Processors, by Antonio Ferrer Sánchez, Carlos Flores and Alejandro Gómez Cadavid, carrying a publication date of 10 July 2026. Quantum Computing Report brought it to wider attention on 17 October 2026. The post is not a peer-reviewed paper, and it neither links nor cites a preprint with the full methods. The figures below are the authors' own, read from the post.
How Kipu Quantum turned each molecule into a program for 200 atoms on QuEra's Aquila
Aquila is an analog machine. It holds single atoms in place with focused laser beams and drives them with smooth light pulses into a highly excited Rydberg state; when one atom is excited, its close neighbors respond, and this interaction shapes how the whole array evolves. QuEra's product page states that Aquila has been publicly available through Amazon Braket since 2022, with a minimum atom spacing of 4 micrometers.
Kipu Quantum's method, which it calls quantum feature extraction, writes the description of each molecule into the settings of that evolution. The team assigned each molecule its own set of driving schedules and atom spacings, laid out 200 atoms on a grid of 14 by 15 positions a few micrometers apart, and let the system evolve for a little under four microseconds before reading out every atom. The authors write that they reached the machine through their own Kipu Quantum Hub platform.
They submitted the training set one molecule at a time, with 500 independent measurements per molecule, and processed 122 molecules in total, roughly two thirds labeled non-toxic and one third toxic. From the strings of excited and unexcited atoms they built about 1,000 features per molecule. A conventional classifier then made the final toxic or non-toxic decision. The quantum processor never classified anything itself; it acted as a transformation step in front of standard software.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The method has run on real quantum hardware as a proof of concept on 122 molecules from one public research dataset, and the sources reviewed here report no test on a pharmaceutical company's compound library, no independent replication and no use in a safety assessment.
Which dataset Kipu used: 171 molecules aimed at the circadian clock protein CRY1, from Koç University
The data come from the Toxicity dataset in the UCI Machine Learning Repository, donated in 2022 by Şeref Gül and Fatih Rahim of Koç University in Istanbul. It holds 171 small molecules, each described by 1,203 standard molecular descriptors, of which 56 are labeled toxic.
The molecules were designed for a specific biological purpose. In a 2021 paper in Scientific Reports, Gül, Rahim, Ibrahim Halil Kavakli and colleagues used structure-based screening to select 171 candidate molecules aimed at functional domains of the core clock protein CRY1, part of the machinery that sets the body's 24-hour circadian rhythm. They tested the molecules in human U2OS cells and found 115 of them non-toxic. Their own decision-tree classifier, using 13 selected descriptors, predicted toxicity with a mean accuracy of 79.53 percent under tenfold cross-validation.
That last figure matters for reading the new result. Kipu's classical baseline reached a raw accuracy of about 0.74, below the 79.5 percent the original authors reported for their own classical model. The two numbers come from different protocols, different splits and a different number of molecules, so they cannot be compared directly. The gap is a reason to ask for matched comparisons, in which the classical and quantum pipelines are tuned with the same care on the same molecules.
What the scores show: balanced accuracy from 0.55 to 0.70 and F1 from about 0.30 to 0.60
Because only about a third of the molecules are toxic, a model can score well on plain accuracy by calling most compounds safe. The authors therefore reported four measures: plain accuracy, balanced accuracy, the F1 score and the area under the ROC curve. They trained XGBoost, CatBoost, a random forest, a support vector machine and Google Research's TabFM tabular model on classical features, on quantum features and on both combined, and averaged every result over 100 random splits of the data.
According to the post, plain accuracy rose only modestly, to close to 0.78 for the strongest models. The larger change was in the measures that punish a model for missing toxic compounds. Balanced accuracy moved from about 0.55 with classical features to about 0.70 with quantum features, and the F1 score roughly doubled, from about 0.30 to about 0.60. A support vector machine on quantum features alone gave the highest area under the curve in the study, about 0.81. The authors write that they applied feature selection to both feature sets so that the strongest classical model was compared with the strongest quantum one.
Several limits apply. The post reports 122 molecules processed on the hardware, so each test split holds only a few dozen compounds and a handful of toxic ones, and differences of this size can move with small changes in method. The post does not report whether a classical transformation of comparable size, such as a random nonlinear projection producing a thousand features, would give similar gains. Such a control would help show how much of the improvement comes from the quantum dynamics and how much from the larger feature space. The Monitor met the same question in the SpinQ and DIVBIO study of GLP-1 receptor potency prediction, where a careful classical control shaped how a small quantum gain should be read.
What drug-safety scientists and hospital pharmacists can take from the October 2026 report
Toxicity prediction from structure is a routine early filter in drug discovery, and models that catch more of the toxic minority would let chemists drop risky compounds before animal or human testing. A method that improves balanced accuracy is aimed at the right weakness.
For clinicians and hospital pharmacists, the report changes nothing in practice. The compounds are research molecules aimed at a clock protein, the toxicity label comes from a cell assay, and no medicine on the market was assessed. Toxicity screening in drug development will continue to rest on laboratory assays, animal studies where required, clinical trials and classical prediction models.
For research leaders following quantum methods in chemistry, three developments would make the result more convincing: a full paper with methods and code that others can rerun; tests on larger toxicity benchmarks with thousands of compounds; and a comparison against strong classical feature transformations of the same size. Kipu Quantum writes that it will soon report further findings.
Sources
Primary source: Antonio Ferrer Sánchez, Carlos Flores and Alejandro Gómez Cadavid, Kipu Quantum, technical blog post on neutral-atom quantum feature extraction for molecular toxicity, dated 10 July 2026. Also drawn on: QuEra Computing's Aquila product page, the UCI Machine Learning Repository record of the Toxicity dataset by Şeref Gül and Fatih Rahim, the 2021 Scientific Reports paper by Gül, Rahim, Kavakli and colleagues, Quantum Computing Report's coverage of 17 October 2026, and the Monitor's earlier coverage. The readiness rating is the Monitor's editorial assessment.