Pillar C · State as of 2026-10-11

How do you read a “12,635-atom protein on a quantum computer” claim?

As of October 11, 2026, the largest protein modeled with help from a quantum computer is a 12,635-atom trypsin–ligand complex in water, reported by Cleveland Clinic, RIKEN and IBM (arXiv:2605.01138, May 2026) and named a finalist for the 2026 ACM Gordon Bell Prize on September 9, 2026. Read past the headline: the quantum processors touched at most 94 qubits’ worth of the most entangled fragments, the «210 times improvement in accuracy» covers «a specific step of the workflow» measured against the team’s own previous quantum pipeline, and IBM’s own write-up states that the method «does not yet outperform the best classical approaches» (IBM Quantum blog, May 5, 2026) — while fully classical fragment chemistry ran a ≈100,000-atom system at MP2 level in about 2 hours on Fugaku in February 2025 (J. Comput. Chem. 46, e70052). Rosetta Quantum reads an atom count the way it reads any claim: by asking what level of theory, what quantity against what reference, and who — quantum or classical — did which part of the work.

→ Leer en español

State as of: 2026-10-11

Status as of: October 11, 2026. Every number in this post carries its source and date beside it. Six earlier Rosetta Quantum posts are cross-linked below and not re-litigated here.

What exactly does the claim say?

That a hybrid quantum–classical pipeline computed electronic-structure energies for the «largest biologically meaningful molecules ever modeled with quantum computers» (Cleveland Clinic release, September 9, 2026): T4-lysozyme with a bound ligand, 11,608 atoms, and trypsin with a bound ligand, 12,635 atoms, both in explicit water (IBM Quantum blog, May 5, 2026; the preprint is arXiv:2605.01138, with «updated results» published in September 2026 per the same release). On September 9, 2026 the team was named a finalist for the 2026 ACM Gordon Bell Prize — a prize for «outstanding achievement in high-performance computing», not a quantum-advantage award — and the winner will be announced at SC26 in Chicago, November 15–20, 2026 (same release). The claim is real, dated and sourced, and this post is not a debunk: it is a worked example of our paper-reading protocol applied to one number — the atom count — because that is the number the headlines carry and the one easiest to misread.

Which part of the 12,635 atoms did the quantum processors actually touch?

The most entangled fragments — up to 94 qubits at a time, on two 156-qubit IBM Quantum Heron r2 processors (ibm_cleveland at Cleveland Clinic, ibm_kobe at RIKEN), running «nearly 6,000 quantum operations» on those parts of the problem (Cleveland Clinic release), across 9,200 circuits and 1.3 billion measurement outcomes over 100+ hours of execution (IBM Quantum blog, May 5, 2026). Everything else in the pipeline is classical, and the claimants say so: wave-function-based embedding splits each complex into clusters; an MP2 step builds each fragment’s environment, made affordable by restricting it to local spheres of 7–10 ångströms around each atom (QCR, May 5, 2026); the simpler clusters are solved classically; and the quantum samples are reassembled by configuration recovery and subspace diagonalization on Fugaku and the GPU-accelerated Miyabi-G — about 30,000 orbitals in the largest run (IBM Quantum blog).

Step of the pipeline	Runs on	Scale	Source
Fragmentation into clusters (wave-function-based embedding)	classical	the full 11,608- and 12,635-atom complexes	IBM Quantum blog (2026-05-05)
MP2 bath construction	classical	local spheres of 7–10 Å around each atom	QCR (2026-05-05)
Most entangled clusters (SQD + TrimSQD)	quantum — two 156-qubit Heron r2 (ibm_cleveland, ibm_kobe)	up to 94 qubits · 9,200 circuits · 1.3 billion measurements · 100+ hours	IBM Quantum blog (2026-05-05)
Configuration recovery + diagonalization	classical — Fugaku and Miyabi-G	≈30,000 orbitals in the largest run	IBM Quantum blog (2026-05-05)
The headline	press materials	«12,635 atoms»	Cleveland Clinic release (2026-09-09)

That split is not a flaw — it is the design, and it raises the question our hybrid-pipelines post taught to ask of every workflow labeled quantum: which side does the work? Here the answer is printed by the claimants themselves. The atom count measures how far the classical embedding reaches; the qubit count measures how much of the problem the quantum processors saw. A reader who compares «12,635 atoms» against any purely classical atom count is comparing two classical fragmentation schemes — one of which has a quantum step inside.

What is the 210x accuracy figure measured against?

Against the team’s own previous quantum-centric pipeline — not against classical chemistry. IBM’s technical blog states the result «represents a 210-times improvement in accuracy from previous state-of-the-art QCSC approaches in a specific step of the workflow», and the same page states the method «does not yet outperform the best classical approaches» (IBM Quantum blog, May 5, 2026). The Cleveland Clinic release prints «210 times improvement in accuracy» with no baseline attached (September 9, 2026) — the denominator lives in the technical post, not in the press materials, which is exactly where our reading protocol says to look. The same discipline applies to the size figure: the release says the team «scaled their method roughly 40 times», and the baseline is the team’s own 303-atom Trp-cage simulation — the first electronic-structure calculation of a protein with a quantum-centric workflow, arXiv:2512.17130, reported with accuracy «competitive with high-level classical benchmarks such as MP2 and CCSD» (QCR, March 26, 2026); 12,635 ÷ 303 ≈ 41.7 is our own declared arithmetic. Competitive, not superior; one workflow step, not the whole simulation; their own prior pipeline, not the classical state of the art. None of this is hidden — all of it sits one click below the headline, printed by the claimants.

How big do purely classical methods already go?

About eight times bigger, at a correlated level of theory, in about two hours. In February 2025, a fully classical fragment-molecular-orbital calculation (FMO-MP2 with the 6-31G* basis set, frozen core) ran a SARS-CoV-2 spike-protein droplet model of about 100,000 atoms in «about 2 h on 8 racks» — 3,072 nodes — of Fugaku (J. Comput. Chem. 46, e70052, February 2, 2025). That is the same Fugaku that serves as the classical half of the quantum-centric pipeline, and 100,000 ÷ 12,635 ≈ 7.9 is our declared arithmetic. The point is not that classical wins a race: the two calculations compute different quantities, at different levels of theory, on different systems — which is precisely why their sizes and times cannot be ranked against each other. The point is that an atom count, by itself, has never been the quantum frontier: fragmentation methods have put correlated quantum chemistry on five- and six-digit atom counts for years, classically. What the May 2026 result demonstrates is that a quantum processor can now sit inside such a pipeline at meaningful scale — a real engineering milestone, and one its claimants do not present as quantum advantage. The feasibility verdicts of our molecular-simulation post and our materials post are unchanged by it.

What three questions make two simulation claims comparable?

First: what level of theory, in what basis set? An energy at MP2 in a modest basis is not an energy at CCSD(T) in a large one, and a size or time attached to one says nothing about the other. The materials accessible to this run state MP2 for the bath-construction step and do not state the level or basis of the headline runs — which is why the scoreboard above prints «n/s». Second: what quantity, against what reference? These runs estimate ground-state energies of fragments reassembled into complexes; the accuracy figure in circulation is measured against the team’s own earlier pipeline — not against an exact result, an experiment, or a classical method on the same instance. Third: who did which part, at what cost? One hundred-plus hours of quantum execution on the most entangled fragments, versus two classical hours for an eight-times-bigger system computing something else, do not divide into a meaningful ratio — and Rosetta Quantum will not print one. This is a case where the ledger’s margin unit R cannot be computed, so the ledger reports asymmetrically and says so: what is measured (atoms, orbitals, qubits, circuits, hours) and what is missing (a same-instance, same-quantity classical baseline at a matched budget). Any future claim that this class of pipeline beats classical chemistry will have to supply exactly that baseline.

What would move this claim up the ledger?

A published same-instance comparison: the same complex, the same target quantity, the best available classical method at a declared budget, and the margin printed with its interval. The claimants’ own forward-looking statements will be read when they come due: lead author Kenneth Merz expects that quantum computing «may soon surpass the best purely classical approaches in chemical accuracy» (QCR, May 5, 2026) — a future-tense expectation, logged as such. The Gordon Bell decision will land at SC26 on November 15–20, 2026, and will add peer scrutiny to the supercomputing claim — the prize judges HPC achievement, not quantum advantage, so its outcome will move the engineering row, not the advantage row: the counter of advantage-backed commercial workloads in our usage map read 0 on October 6, 2026, and nothing in today’s reading moves it. Rosetta will re-read this claim when the SC26 result is published.

What we don’t know

We don’t know the level of theory or the basis set of the headline runs: the arXiv preprint (2605.01138) was not accessible from this run’s environment, so every quote here comes from IBM’s technical blog, Cleveland Clinic’s releases and dated secondary analyses (PostQuantum, May 6, 2026) — a sourcing limitation we declare rather than hide. We don’t know which «specific step of the workflow» the 210x figure measures, or in what metric: no material accessible to this run defines it. We don’t know what changed between the May preprint and the «updated results» the September release mentions. We don’t know the end-to-end wall-clock of the full pipeline: the 100+ hours attach to the 9,200-circuit quantum workload, the accessible materials do not say what that figure includes, and no total pipeline time or classical core-hours figure is published — so the quantum fraction of the whole cannot be computed. And we don’t know whether a same-instance classical baseline exists unpublished: if one does, that number — not the atom count — will decide where this claim sits on the ledger.