IonQ and Kipu Quantum's 64-Qubit Protein-Folding Paper Took Third Place at IEEE Quantum Week 2026: Six Peptides, 46 to 61 Qubits, Four of Six Reference Energies Reached After Classical Post-Processing
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · September 16, 2026.
A Kipu Quantum and IonQ paper on lattice protein folding took third place in the hybrid case-studies track at IEEE Quantum Week 2026 in Toronto, IonQ announced on 15 September 2026, and the release says the work "was used to improve computational speed and accuracy in drug discovery." The paper, on arXiv since 29 April 2026, reports something narrower: a hardware benchmark on six short peptides that reached the reference energy of a conventional solver in four cases, and only with the help of a classical consensus step.
The experiment ran six peptides of 14 to 16 residues through a 64-qubit trapped-ion processor, a barium development system IonQ describes as similar to its forthcoming Tempo line, folding each on a coarse lattice with no side chains using 46 to 61 qubits. The quantum circuits produced samples that mostly violated the backbone geometry, and it took a classical post-processing step, a consensus vote over the two thousand lowest-energy samples combined with a pool of feasible backbones, to reach the reference energy in four of the six cases; a random-start version of the same pipeline reached it in one. For a hospital buyer or a clinical pharmacologist that is the whole result: a hardware run on a benchmark, at a scale that a classical genetic algorithm solves to convergence, with the quantum contribution visible in the statistics of one part of the problem.
IonQ's 15 September 2026 release and the IEEE Quantum Week award list
The IonQ release counts 857 submissions and 27 Best Paper honors at the conference, four of which went to papers with IonQ authors. Three of the four concern AI fine-tuning, distributed optimization with Oak Ridge National Laboratory and NVIDIA, and linear algebra with Synopsys. The fourth is the protein-folding paper with Kipu Quantum, which the release assigns to drug discovery. The conference's award list gives the paper third place in the Quantum End-to-End Hybrid Case Studies track, with the paper presented on Monday 14 September and the track's awards presented on Tuesday 15 September. Five of the fourteen authors are at Kipu Quantum in Berlin, one, Sebastián Romero, at ICMM-CSIC and the Autonomous University of Madrid, and eight at IonQ, among them Martin Roetteler, IonQ's vice president of quantum applications research, who is quoted in the release. The award recognizes a conference paper; it does not evaluate a drug, a target or a clinical claim, and the paper makes none.
What the 29 April 2026 paper ran: six lattice peptides on a 64-qubit barium system
The arXiv paper uses the tetrahedral lattice model that Robert, Barkoutsos, Woerner and Tavernelli published in npj Quantum Information in 2021: every amino acid becomes one bead on a tetrahedral lattice, side chains are ignored, each turn between consecutive beads takes two qubits, and a second set of qubits flags which pairs of beads separated by five or more positions come into contact. Contacts are scored with the Miyazawa-Jernigan statistical potential, and a penalty term forbids the chain from doubling back on itself. A 14-residue chain needs 46 qubits (21 for geometry, 25 for contacts), a 15-residue chain 53, and a 16-residue chain 61. The six sequences are real molecules chosen as benchmarks: mastoparan I, a wasp-venom peptide; a designed beta-hairpin; CM15, a synthetic cecropin-melittin hybrid; a fragment of tissue transglutaminase; MOTS-c, a mitochondrial peptide involved in metabolism; and VG16KRKP, a cationic antimicrobial peptide.
The algorithm is bias-field digitized counterdiabatic quantum optimization, BF-DCQO, which Kipu Quantum has used before. It has no variational loop: the circuit parameters come from an analytic formula, and after each round the lowest-energy measurements set longitudinal bias fields for the next round. The authors ran ten rounds per sequence, eight for MOTS-c because of hardware availability, with 5,000 shots per round, and pruned the circuits so that the number of entangling gates stayed near one thousand. The processor is the 64-qubit barium development system with all-to-all connectivity, so the dense pairwise couplings of the contact term can be implemented without swap gates. The same group's June 2025 paper reached 12 amino acids on 33 qubits of an IonQ system; the new work extends that to 16 residues and 61 qubits, and the paper describes itself as the largest trapped-ion demonstration of lattice protein folding to date.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The optimization ran on quantum hardware that exists, on six short benchmark peptides whose folded lattice energies were already known from a classical solver, and it needed a classical post-processing step to reach those energies in four of six cases. Nothing here has been assembled into a drug-design workflow, tested against a therapeutic target, or compared with the structure-prediction tools a pharmaceutical team uses today.
What "four of six reference energies" means, and why the 61-qubit runs reached none
The reference energies come from a classical genetic algorithm run to convergence, so the benchmark is whether the quantum-seeded workflow can match a method that already solves these instances. The raw quantum samples do not: because the circuits do not enforce the no-backtracking constraint and the hardware is noisy, most measured bitstrings describe chains that overlap or fold back on themselves. What the authors show is that the distribution of raw energies shifts downward relative to uniform random bitstrings, by a factor of 2.9 in mean energy for the transglutaminase fragment at 53 qubits, and that the shift concentrates in the contact qubits, which make up the majority of the register.
Turning that shift into a fold takes the consensus pipeline. It selects the 2,000 lowest-energy samples, takes a majority vote on each contact qubit, and scores 200 candidate backbone geometries against the voted contact pattern before a final repair. Seeded with quantum samples, this pipeline reached the reference energy for all four sequences at 46 and 53 qubits; seeded with random bitstrings, it reached it once. Mean energies after post-processing tell the same story, for instance −4.09 against −2.07 for LEPFSGKALCSWSIC, where the reference is −8.46. At 61 qubits the two 16-residue sequences, MOTS-c and VG16KRKP, reached the reference energy in neither case, and the authors attribute this to the pruning threshold they chose to hold the entangling-gate count near one thousand. They also show that a more naive post-processing step, which repairs each sample independently and re-optimizes its contacts from scratch, erases the pattern the quantum rounds had learned. The quantum processor, in their own summary, supplies structured statistical information about contacts, and classical code does the rest.
Where a 16-residue lattice peptide sits against a drug target, and against the 114-qubit and 45-qubit runs
A hospital reader should hold three scales apart. The peptides here are 14 to 16 residues; the benchmark has meaning because the reference energy is computable, and the same property means a classical solver already covers it. Therapeutic proteins and their targets run to hundreds of residues, with side chains, solvent and flexibility that the tetrahedral bead model leaves out by design, and for those the deep-learning predictors set the practical standard, as the paper's own introduction acknowledges. Short peptides are a real drug class, and quantum hardware has been pointed at them before from a different direction: the photonic sampling study that this Monitor's general lane read in July used a boson sampler to seed a generative model of nine-residue peptide-HLA binders, and its open question was population coverage across HLA alleles. The Kipu-IonQ work asks a different question of a similar-sized molecule, whether a lattice fold can be optimized on hardware, and a 16-residue backbone with no side chains answers it as a proxy rather than as a peptide.
The paper places itself among gate-model competitors with care. It cites a 2024 run that reached 20 amino acids on 114 superconducting qubits with a different turn encoding, and a 2025 problem-agnostic approach that reached 26 amino acids on 45 superconducting qubits; the first author of that second paper, Hanna Linn, then at Cleveland Clinic, is now among the IonQ authors of this one. The Kipu-IonQ claim is specific: the largest trapped-ion run, and the largest with the dense 2021 encoding. That is a claim about a hardware platform and an encoding, and it holds regardless of whether trapped ions turn out to matter for pharmaceutical work. Readers who followed QC Ware and IonQ's eight-qubit cytochrome P450nor calculation on Forte on 1 September will notice one difference in favor of this paper: that release left its methods and its classical benchmark unidentified, while the arXiv paper here states its classical reference method, its shot counts, its pruning levels and its per-sequence energies in a table.
How Quentir Reads It
The press release says the paper used counterdiabatic optimization "to improve computational speed and accuracy in drug discovery." The paper reports no speed comparison against a classical solver and no accuracy measure beyond matching a genetic algorithm's converged energy on a lattice; its stated contributions are a first application of BF-DCQO in the 46-to-61-qubit range on trapped ions, evidence that the quantum samples carry contact structure, and a demonstration that post-processing must be designed to keep that structure. Those are contributions to quantum optimization, rewarded in a track built for hybrid case studies. The drug-discovery sentence belongs to the release, and a procurement file that cites this award should attach the paper rather than the release.
The practical detail for anyone following the hardware is the entangling-gate budget. The authors held the entangling-gate count near one thousand, and at that budget the 61-qubit instances failed where the 46- and 53-qubit ones succeeded. A useful follow-up test, whether from this pair on Tempo-class hardware or from another group, would raise that budget on the same six sequences and report three things together: the gate count at which a 16-residue instance reaches its reference energy, the pruning level and noise at which the contact signal disappears, and the same classical genetic algorithm run for the same wall-clock time. Until such a comparison exists, the result stands as a hardware benchmark on six short peptides.
Image: an AI-generated conceptual illustration of a lattice model of a short peptide, ceramic beads threaded on one continuous copper cord in front of a separate tetrahedral scaffold. It depicts no real product, processor or laboratory and no clinical use; the model stands for the side-chain-free encoding in the paper and shows fewer beads than the 14 to 16 residues of the studied sequences.
Sources
Primary source: Alejandro Gomez Cadavid, Pavle Nikačević, Pranav Chandarana, Enrique Solano and Narendra N. Hegade of Kipu Quantum, Sebastián V. Romero of ICMM-CSIC and the Autonomous University of Madrid, with Miguel Angel Lopez-Ruiz, Claudio Girotto, Hanna Linn, Hakan Doga, Evgeny Epifanovsky, Panagiotis Kl. Barkoutsos, Ananth Kaushik and Martin Roetteler of IonQ, "Protein folding on a 64 qubit trapped-ion hardware via counterdiabatic quantum optimization," arXiv 2604.26861, 29 April 2026. Also drawn on: the IEEE Quantum Week 2026 Best Papers list; IonQ's press release of 15 September 2026; Robert, Barkoutsos, Woerner and Tavernelli in npj Quantum Information, 2021; and Romero and colleagues, arXiv 2506.07866, June 2025. The readiness placement is this Monitor's own.