The 12,635-atom quantum chemistry run never gave the quantum chip more than 45 orbitals at once
The biggest piece of the protein simulation that ran on quantum hardware held 66 electrons. Here is how the remaining work was split, and how classical methods compared.
The trypsin-benzamidine complex simulated by Cleveland Clinic, RIKEN and IBM contains 12,635 atoms, 43,858 electrons and 31,795 orbitals. No single quantum computation in the study handled more than 66 of those electrons spread over 45 orbitals. Classical supercomputers did everything else. They cut the protein into pieces, turned quantum measurements into energies and reassembled the results. A fully classical version of the same workflow also produced binding energies comparable to the hybrid version.
The team describes this division of labor in its preprint, "Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing" (arXiv 2605.01138). The preprint is not yet peer reviewed; it was first posted in May 2026 and revised in September. IBM announced the work on May 5, 2026. It said 156-qubit IBM Quantum Heron processors at Cleveland Clinic in the United States and at RIKEN in Japan worked alongside the Fugaku and Miyabi-G supercomputers. Cleveland Clinic says the project is a finalist for the 2026 ACM Gordon Bell Prize, with the winner to be announced at the SC26 conference in Chicago (November 15 to 20, 2026).
Cutting a protein into pieces
An orbital is a region an electron can occupy. An electronic configuration is one pattern of which orbitals hold which electrons. The preprint notes that the number of possible configurations grows combinatorially with system size. For that reason, the exact method, called full configuration interaction (FCI), only works for small molecules.
The team therefore used a quantum embedding method called EWF. Embedding splits a molecule into fragments that can be solved separately and then recombined. In EWF, each fragment is one atom plus a surrounding "bath" of electrons and orbitals that are entangled with it. The trypsin model was built from crystal structure 3PTB and kept 3,135 water molecules within 10 angstroms of the protein. It used STO-3G, a minimal basis set, meaning the smallest standard set of orbitals per atom.
Classical computers did the setup. This included a Hartree-Fock calculation (a mean-field approximation) over the whole system and the construction of each fragment from orbitals within a 7-angstrom sphere around its central atom. Fragments centered on hydrogen atoms with fewer than 13 orbitals were solved exactly with FCI on classical machines. Fragments with 13 or more orbitals went to the quantum processors.
In the baseline trypsin calculation, those fragments fell into two size groups: 13 to 26 orbitals and 27 to 45 orbitals. They received 100,000 and 500,000 measurement shots respectively, using about 29 seconds and 2 minutes 22 seconds of processor time per fragment. In total they took 56 hours on the ibm_cleveland processor. The bound trypsin calculation alone included 403 non-water fragments with 27 or more orbitals.
What the quantum chip did, and what it did not
The quantum step used sample-based quantum diagonalization (SQD). The quantum processor runs a circuit and measures it many times, and each measurement outcome reads out as one electronic configuration. A classical computer then builds a much smaller version of the energy problem using only those sampled configurations. It diagonalizes that problem, which means it finds the lowest energy. A separate IBM preprint describes the core idea the same way: the ground-state problem is solved classically, inside a subspace built from quantum samples.
The diagonalization itself was heavy classical work. In the baseline runs, each fragment job used 1,152 Fugaku nodes. In portability runs on Miyabi-G, each fragment used 8 to 64 GPU nodes, and a third RIKEN system, ROQUO, ran the automated end-to-end version. The team's main new method, TrimSQD, changes this classical post-processing step.
The largest quantum circuit belonged to fragment 3211, which has 66 electrons in 45 orbitals. Its full configuration space holds about 8.27 x 10^20 configurations. The circuit used 94 physical qubits, 60 percent of the 156 on ibm_cleveland. It ran 5,758 two-qubit gates (operations between pairs of qubits) at a two-qubit depth of 246 and measured 90 qubits. The preprint writes each configuration as a string of two bits per spatial orbital, so 45 orbitals map to 90 bits. The two-qubit part of the circuit took about 17 microseconds, which is shorter than the median coherence times (how long qubits hold information) of the qubits used.
Across all calculations, the revised preprint reports 21,006 circuits, more than 239 hours of quantum processor time and 3.0 billion measurement outcomes. IBM's May blog post had reported 9,200 circuits, more than 100 hours and 1.3 billion outcomes. Cleveland Clinic says the team published updated results in September.
Could classical methods alone get the same answer
For this study, the public record largely answers yes. The team ran the whole workflow a second time with a purely classical fragment solver, CCSD (coupled-cluster singles and doubles, a widely used approximation). In the minimal basis, the classical version gave a trypsin binding energy of 28.46 kcal/mol, compared with 35.89 kcal/mol for the hybrid version. The authors call both results comparable but too positive. After the authors added an electrostatic correction, the figures were -130.59 and -131.32 kcal/mol. For the T4-Lysozyme complex (11,608 atoms) in a mixed basis, they were -17.31 and -21.07 kcal/mol.
The accuracy figures themselves come from classical methods. The "210 times" improvement compares how far each result lands from a classical DMRG reference. DMRG (density matrix renormalization group) is a tensor-network method. The gap was 21 milliHartree for the largest fragment of the earlier 303-atom Trp-cage study and 0.1 milliHartree for the largest trypsin fragment (a milliHartree is a small unit of energy). For that 66-electron fragment, TrimSQD landed 0.098 milliHartree above DMRG. Selected configuration interaction, a classical method that picks configurations without a quantum computer, landed 0.046 above. DMRG gave the lowest energy on all three benchmark fragments.
The authors write that they "do not regard the results reported here as basis of a quantum advantage claim." IBM's blog says the method "does not yet outperform the best classical approaches." The authors also plan to add their fast diagonalization code to classical selected-CI software, which they expect could significantly cut its run time.
The record does not show whether the quantum step adds accuracy that classical solvers could not match at the same cost. Answering that would take fragments too large for DMRG or selected CI to converge, with both approaches compared on matched computing budgets. The authors project that future error-corrected machines with 200 to 2,000 logical qubits could handle fragments of 100 to 1,000 orbitals.
What newer preprints are fixing
One line of work notes that sampled configurations generally do not form a set with a well-defined total spin. A proposed fix is to add the missing partner configurations before the classical diagonalization.
One test used IBM's ibm_boston Heron r3 processor, with a manganese complex called oxo-Mn(Salen) and an active space of 32 electrons in 24 orbitals. The system is small enough for a converged DMRG reference. The spin-adapted method returned exact spin states while diagonalizing subspaces two to three orders of magnitude smaller than conventional constructions. The classical heat-bath configuration interaction method was still more accurate in absolute terms, but it needed subspaces of 4 to 15 million configurations. The spin-adapted circuits themselves have not yet run on hardware, and the authors say they would substantially increase circuit depth.
- Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing: quantum chemistry of protein-ligand complexes (abstract), arXiv (preprint, not yet peer reviewed)
- Cleveland Clinic, RIKEN, and IBM Model a 12,635-Atom Protein – the Largest Known to Be Simulated with Quantum Computers, IBM Newsroom
- Researchers use quantum-centric supercomputing to simulate 12,635-atom protein complex, IBM Quantum Blog
- Cleveland Clinic, RIKEN and IBM Team Advance to Finals for 2026 ACM Gordon Bell Prize, Cleveland Clinic Newsroom
- Spin-adapted sample-based quantum diagonalization with randomized time evolution, arXiv (preprint, not yet peer reviewed)