QuEra's Yuval Boger Expects Materials Before Drug Discovery on Fault-Tolerant Quantum Computers: What the Harvard, Anthropic and QuEra Preprint arXiv 2610.22367 Cut From Chemistry Circuits

Quentir Medicine Monitor

Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · November 19, 2026.

Conceptual close-up of a translucent blood vessel with red blood cells drifting through it; a thin ring of gold light circles the vessel wall, and the cell in focus holds a small faceted amber crystal, an image for the iron at the centre of a metalloprotein.

In an interview that R&D World published on 17 November 2026, Yuval Boger, chief commercial officer of the Boston neutral-atom company QuEra Computing, said he does not expect drug discovery to be the first major payoff of fault-tolerant quantum computers. He expects strongly correlated materials to come first, with molecular chemistry, including metalloprotein active sites, following on QuEra's next-generation system. Much of the interview concerned a preprint in which a large language model designed quantum circuits for chemistry and other tasks that need fewer gates and fewer qubits than the designs the authors compared them against.

For pharmaceutical and hospital research teams the two pieces belong together. The preprint lowers the estimated cost of several building blocks that a quantum chemistry calculation needs, and the interview sets out which problems one of the companies building the hardware expects those savings to reach first, and on what timetable. Neither involves a run on quantum hardware. The preprint's circuits are written for fault-tolerant hardware, and the interview's timeline concerns machines that QuEra plans to deliver from 2028.

What Gu, Kemeny, Yelin and Lukin reported in arXiv 2610.22367 on 15 October 2026

The Harvard, Anthropic and QuEra preprint on arXiv, posted on 15 October 2026, lists Andi Gu of Harvard's Department of Physics and the Harvard Quantum Initiative, who is also part of the Anthropic STEM Fellows Program; Alek H. Kemeny of Anthropic; Susanne F. Yelin of Harvard; and Mikhail D. Lukin of Harvard and QuEra Computing. It has not been peer reviewed. The authors built an evolutionary search in which a language model, Anthropic's Claude Mythos 5 running inside the open-source OpenEvolve framework, proposes programs that generate circuits, and an independent evaluator checks each candidate for correctness and scores its cost. For each of eight tasks the model receives a fixed, hand-written specification and a shared circuit-building interface. The tasks span numerical function evaluation, chemistry, optimization, quantum simulation and number theory.

The cost measure is the one that dominates estimates for error-corrected machines: the number of Toffoli gates, which are expensive to carry out fault-tolerantly, and the number of qubits a circuit occupies. In seven of the eight tasks the discovered circuits reduced both counts below the comparison baselines. For one decoding task the authors note that no earlier reversible compilation existed in the literature, so the baseline there was a naive reversible implementation. In the eighth, modular inversion over a prime field, a step in cryptographic arithmetic, the search produced a new reversible architecture that substantially improves the trade-off between Toffoli gates and qubits.

How the iron-sulfur and Coulomb-potential circuits shrank in the preprint's resource estimates

Two of the tasks bear directly on chemistry. The first is the Coulomb potential, the electrostatic interaction between charged particles that a simulation of electrons and nuclei evaluates again and again. The reference circuit used 6,355 Toffoli gates and 252 qubits; the discovered circuit used 1,140 Toffoli gates and 142 qubits, a 9.9-fold improvement in the product of the two counts. The second is a component of a chemistry simulation for the iron-sulfur cluster Fe2S2, a standard benchmark. Its Toffoli count fell from about 55,000 to about 15,000 and its qubit count from 139 to 126, a 4.1-fold improvement in the same product.

Boger was careful about what these figures show. "These are verified circuit designs and resource estimates for building blocks, not hardware runs and not full applications," he told R&D World. A full simulation gains only as much as the improved building block contributes to its total cost, so each application has to be checked on its own. A fault-tolerant machine has a fixed budget of logical qubits and reliable operations, and a cheaper building block lets a larger molecule, or a larger active space, the set of electrons and orbitals a calculation treats at full accuracy, fit within that budget.

Quantum pillar: computing. Technology readiness: TRL 2 of 9. In Quentir's assessment, the work sits at the concept stage: the circuits exist as verified designs and cost estimates on classical computers, none has run on a quantum processor, and no drug, protein or clinical workflow has used them.

Why Boger puts strongly correlated materials ahead of drug discovery, and where metalloproteins fit

"People often expect drug discovery to be the first big win. I don't," Boger said. His reason is the strength of the classical tools that pharmaceutical chemists already use. Density functional theory and machine-learned force fields are very good, in his words, and AI makes them better every year. He expects error correction to pay off first on strongly correlated materials, where different classical methods give different answers, and he gave Fermi-Hubbard lattice models of up to 10 by 10 sites as an example. He added that AI helps classical methods as much as quantum ones, so the bar a quantum computer has to clear keeps rising.

Molecular chemistry, in Boger's account, comes next. He placed ground-state energies of strongly correlated molecules with small-to-medium active spaces, such as metalloprotein active sites and transition-metal chemistry, on QuEra's next-generation system, which targets more than 1,000 logical qubits at an error rate of one in a billion, about a billion reliable operations, in 2028 or 2029. He called that a roadmap target and declined to present it as a delivery date. He cited the iron-molybdenum cofactor of the enzyme nitrogenase, known as FeMoco, as evidence that costs fall: its estimated cost dropped from about 50 trillion Toffoli gates to about 340 million in eight years of hand-designed work, and he said the preprint suggests AI can shorten that path.

Boger also described the work closest to medicine. In 2025, Merck, Amgen, Deloitte and QuEra published a study that used quantum reservoir computing on more than 100 qubits to predict molecular properties, with a classical model learning from features the quantum system produced. Two of the six finalists in Wellcome Leap's Quantum for Bio Phase 3 run on QuEra hardware: one studies covalent inhibitors for myotonic dystrophy with Nottingham and Phasecraft, and the other binding affinity and NMR estimation with Harvard and MIT. He expects quantum data for drug discovery to take the form of a small set of very accurate energies, each with an error bar, for the molecules where classical methods are least reliable, used to train or correct classical models. He said this has not been demonstrated and requires fault-tolerant hardware.

What QuEra's 256-logical-qubit Libra system, planned for 2028, means for pharmaceutical and hospital research

QuEra plans to make Libra, designed for 256 logical qubits, a logical error rate of one in a million and about a million reliable operations, available in 2028 on Amazon Braket and at high-performance computing centers. Until Libra is available, Boger said, QuEra can test the AI-designed circuits in emulation, and he said AI has not pulled the Libra date forward. A pharmaceutical estimate gives a sense of the distance. In October, GSK's Anthony Rowe put useful quantum chemistry for problems beyond conventional computers at about 2,000 logical qubits around 2030, roughly eight times Libra's design. Cheaper building blocks of the kind arXiv 2610.22367 reports could lower that requirement for particular calculations. No reduction in Rowe's estimate has been demonstrated: that would require a full resource estimate for a specific molecule, and neither the preprint nor the interview provides one.

QuEra's engineering work with the same model family follows a comparable discipline of human checks. In August the company reported that a Claude-built controller restored a laser's frequency lock in 695 of 700 testbed trials. Boger told R&D World that the five failures traced to a single fault in the test rig, that the controller caught all 43 spontaneous mode hops, and that it now runs as a standalone script engineers can read line by line, with no AI involved at run time and hardware interlocks the model cannot override. For the metalloprotein and transition-metal problems that he links to medicine, his own timeline points to the 1,000-logical-qubit system, which remains a roadmap target.

Sources

Primary source: preprint by Andi Gu, Alek H. Kemeny, Susanne F. Yelin and Mikhail D. Lukin (Harvard University, Anthropic and QuEra Computing), arXiv, 15 October 2026. Also drawn on: R&D World's interview with QuEra's Yuval Boger (17 November 2026), The Quantum Insider's report on QuEra's laser controller (28 August 2026) and earlier Quentir coverage.

  1. Harvard, Anthropic and QuEra preprint on arXiv
  2. Claude-built controller restored a laser's frequency lock in 695 of 700 testbed trials
  3. R&D World
Previous
Previous

Tohoku University Researchers Use Quantum Annealing to Choose Which MRI Lines a Knee Scan Acquires: What arXiv 2611.25625 Tested on fastMRI Images, 19 November 2026

Next
Next

ETH Zurich and Copenhagen Researchers Compute the Energy of an Arginine-Tryptophan Protein Contact on Quantinuum's Helios, arXiv 2611.22834, 17 November 2026