The Minimal-Basis Run Gave T4-Lysozyme the Wrong Binding Sign: What Cleveland Clinic, RIKEN and IBM Changed to Correct It
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · September 9, 2026.
On 9 September 2026 Cleveland Clinic announced that a team from Cleveland Clinic, the RIKEN Center for Computational Science and IBM had reached the finals of the 2026 ACM Gordon Bell Prize, for quantum-classical electronic-structure calculations on protein-ligand complexes of 11,608 and 12,635 atoms. Four days earlier the preprint behind that entry gained a second version, and buried in it is a result that deserves more attention than the atom count.
When the team computed the binding energy of a well-studied protein-ligand pair in a minimal basis set, the answer came out too positive, which is to say it pointed the wrong way for a ligand known to bind. Recomputing it with richer orbitals in the binding region and a tighter bath threshold together corrects the sign of the binding energy, at about ten times more in total node-hours. The revision also reports a first end-to-end wall-clock time, 62.4 hours from fragment generation to energy reconstruction, for a bound-complex ground-state run.
What the 5 September Revision Added to arXiv:2605.01138
The paper is "Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing: quantum chemistry of protein-ligand complexes", by Kenneth M. Merz Jr., Akhil Shajan, Danil Kaliakin and Fangchun Liang of Cleveland Clinic, with colleagues at the RIKEN Center for Computational Science and IBM. Version 1 was submitted on 1 May 2026 and version 2 on 5 September 2026.
The second version is a larger study, and reading its headline totals as the cost of the first version repeated would be a mistake. Version 2 reports six calculations. Two are the baseline binding energies for both systems in a minimal STO-3G basis. Two more are portability runs that reuse the same quantum samples and move the classical diagonalization to a different supercomputer. The fifth is a refinement that recomputes one binding energy in a mixed basis. The sixth measures time to solution using an automated workflow. The totals in the abstract, 21,006 circuits over more than 239 hours of quantum processor time and 3.0 billion measurement outcomes, cover all six. Version 1 reported 9,200 circuits over more than 100 hours for its own smaller set, and it already produced binding energies.
The two systems are standard benchmarks. System I is trypsin with benzamidine, 12,635 atoms, built from crystal coordinates in PDB entry 3PTB. System II is T4-lysozyme with n-butylbenzene, 11,608 atoms and 2,986 retained water molecules, from PDB entry 4W57, a hydrophobic cavity long used to study how a small ligand sits in a nonpolar pocket. Both keep solvent within 10 angstroms of the protein boundary.
Why the Sign of a Binding Energy Is the Result That Matters
Under the paper's electronic-energy convention, a positive result places the computed bound complex above the separated components; it is not a thermodynamic binding-free-energy measurement. For T4-lysozyme with n-butylbenzene, a pairing chosen precisely because the ligand binds well, that answer is wrong on its face. The paper states it plainly: in the minimal-basis calculations the two methods it compares give binding energies that are comparable to each other and too positive, even though the individual fragment solutions are accurate.
That last clause is the interesting part. The fragments were computed correctly and the assembled answer still pointed the wrong way, which locates the error in the description of the chemistry rather than in the quantum sampling. STO-3G is the smallest usable set of building blocks for describing where electrons sit, and a set that small is a poor instrument for the dispersion interaction that holds a hydrocarbon in a nonpolar cavity. The paper does not run the experiment that would isolate the basis as the single cause, so treat that as the readable explanation rather than a demonstrated one.
The fix is targeted rather than global, and it has two parts. The team put richer 3-21G orbitals on the ligand and on every residue with at least one atom within 3 angstroms of it, kept STO-3G everywhere else, and tightened the bath threshold that decides how much of the surrounding protein each fragment sees. Both changes were made together, and the paper reports the combined refinement as what corrects the sign, at roughly ten times the total node-hours of the minimal-basis run. It does not separate the contribution of the basis from that of the threshold, so no reader should credit either one alone.
Quantum pillar: simulation. Technology readiness: TRL 4 of 9. The workflow was assembled and run under laboratory conditions on real quantum processors and real supercomputers, and its fragment energies were checked against an established classical reference, which is rung four on the shared ladder both Evidence Registers use. This Monitor holds it below rung five because rung five asks for testing on the kind of data the real job involves, and the real job in drug discovery is ranking a series of candidate compounds against measured binding affinities. Two benchmark complexes, one of which needed a combined basis and bath-threshold refinement before its binding energy pointed the right way, with no comparison to experimental affinity data, does not reach that bar. The rung moves when this workflow is applied to a compound series and its predicted rankings are compared with laboratory measurements.
What 94 Qubits Computed Inside an 11,608-Atom Complex
The atom count describes the biological scene, not the quantum workload. A quantum embedding step cuts the complex into fragments; the quantum processors sample electronic configurations for those fragments and nothing else; the supercomputers turn those samples into fragment wavefunctions and reassemble them. This Monitor set out that division of labor when the first version appeared, and version 2 leaves it intact. Sampling ran on two 156-qubit Heron r2 processors, ibm_cleveland and ibm_kobe, using up to 94 qubits on one and 72 on the other, which the paper puts at 60 and 46 percent of the available qubits.
The quantum time is substantial and it is only part of the bill. The paper reports 56 hours of processor time sustained on one machine for the first baseline calculation, and 44 and 47 hours across both machines for the second baseline and the mixed-basis refinement. The classical side is larger still, and the automated run on RIKEN's ROQUO system, introduced in June 2026, used 7,620 GPU node-hours at an average of 122.1 nodes.
Set against a drug program, those figures describe a single answer, not a campaign. Lead optimization evaluates hundreds of analogues and repeats the exercise as a series evolves. The paper offers no per-compound scaling, no commercial access rate and no estimate of what a series would cost, so the honest reading is that one binding energy for one complex currently consumes days of dedicated time on two named quantum processors and a national supercomputer. Capacity of that kind is itself now an object of industrial policy: in early September 2026 the US Commerce Department finalized $100 million awards with equity terms for three quantum manufacturers.
Sixty-Two Hours End to End, and What It Does Not Yet Include
The sixth calculation is the one with the clearest operational meaning. Calculations one through five were coordinated by hand across several machines. The last one ran through an automated framework spanning both quantum processors and ROQUO, and produced a mixed-basis ground-state energy for the bound T4-lysozyme complex in 62.4 hours of wall-clock time, measured from fragment generation through to energy reconstruction. Removing manual data transfers between quantum and classical systems is a real engineering result, and it is the first figure in this line of work that a planner can put in a schedule.
Two limits sit next to it. That 62.4 hours covers the bound state only, so a binding energy, which is a difference between the bound complex and its separated parts, requires further calculations. And the accuracy comparison moves in an awkward direction: against a classical reference, the largest fragment calculation deviates by 0.16 millihartree in the minimal basis and 2.23 millihartree in the mixed basis. The mixed basis fixes the chemistry of the assembled binding energy while the individual fragments agree less closely with the reference, which is a distinction a reader should keep hold of.
Six Questions a Pharmaceutical Buyer Should Ask
How much of the 62.4 hours was quantum processor time, and how much was classical compute, queueing and orchestration? What would a full binding energy in the mixed basis cost end to end, given that the automated run covered the bound state alone? How was the mixed-basis region chosen, and how sensitive is the corrected sign to that 3-angstrom cutoff? Have any of these predicted binding energies been compared with experimentally measured affinities for the same complexes, and with what error? What does the same calculation cost using established classical methods at comparable accuracy, since that comparison decides whether a quantum processor earns its place? And how would the cost scale across fifty analogues of one scaffold, at current commercial access rates?
How Quentir Reads It
This is careful work, and the finalist slot reflects a genuine engineering achievement in orchestrating processors, GPUs and quantum hardware on one problem. The candid reporting of a wrong-signed minimal-basis result, and of what it took to correct it, is worth more to a reader deciding whether to invest than the headline atom count. So is the first honest end-to-end wall-clock figure. The winner is announced at SC26 in Chicago, 15 to 20 November 2026.
The commercial reading stays restrained. A method that needed a targeted refinement of both its basis and its bath threshold, at roughly ten times the compute, before one benchmark complex gave a binding energy of the right sign is a method still establishing what it can be trusted to predict. The figure a drug program needs is a validated ranking of candidate compounds against laboratory measurement, and this paper does not report one. This Monitor places the workflow at rung four, treats the mixed-basis binding prediction as a first demonstration rather than a validated capability, and will revisit the entry when predictions from this pipeline are compared with measured affinities across a compound series.
Sources
Primary source: Kenneth M. Merz Jr., Akhil Shajan, Danil Kaliakin and Fangchun Liang of Cleveland Clinic, with Yuichi Otsuka, Tomonori Shirakawa, Lukas Broers, Han Xu, Miwako Tsuji, Mitsuhisa Sato and Seiji Yunoki of the RIKEN Center for Computational Science and Ryo Wakizaka, Yukio Kawashima, Jun Doi, Hitomi Takahashi, Toshinari Itoko, Hiroshi Horii, Thaddeus Pellegrini, Javier Robledo Moreno, Kevin J. Sung, Ella Fejer, Robert Walkup, Seetharami Seelam and Mario Motta of IBM, "Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing: quantum chemistry of protein-ligand complexes," arXiv:2605.01138, submitted 1 May 2026 and revised 5 September 2026. The full text of version 2, read directly rather than from the abstract, supplies the quantitative and technical statements attributed to the paper below; the explanatory passages and the commercial and procurement context in the article are this Monitor's synthesis, not statements from the preprint: the summary of six calculations in Table IV, organized as two baseline binding energies in the STO-3G basis, two portability runs reusing the same quantum samples with diagonalization moved to Miyabi-G, one refinement recomputing the binding energy of system II in a mixed basis, and one time-to-solution run on the automated workflow; the identification of system I as trypsin with benzamidine at 12,635 atoms from PDB entry 3PTB and system II as T4-lysozyme with n-butylbenzene at 11,608 atoms with 2,986 retained water molecules from PDB entry 4W57, both with solvent kept within 10 angstroms of the protein boundary; the statement that in the minimal-basis calculations the classical and quantum-classical fragment solvers yield binding energies that are comparable but too positive although the fragment solutions themselves are accurate; the definition of the mixed basis as 3-21G orbitals on the ligand and on residues with at least one atom within 3 angstroms of it and STO-3G on the remainder, together with the tightened bath threshold; the finding that the combined refinement, the mixed basis together with the tighter bath threshold, corrects the sign of the binding energy at approximately ten times the total node-hours, the paper making no attempt to isolate which of the two changes is responsible; the quantum sampling on two 156-qubit Heron r2 processors, ibm_cleveland and ibm_kobe, using up to 94 and 72 qubits, described as 60 and 46 percent of available capacity, sustained for 56 hours in calculation A and 44 and 47 hours in calculations B and E; the fragment-level deviations against a classical reference of 0.16 millihartree in the STO-3G basis and 2.23 millihartree in the mixed basis; and the end-to-end wall-clock time of 62.4 hours for the automated mixed-basis ground-state calculation of the bound complex only, which is not a complete binding-energy calculation, from fragment generation to energy reconstruction, with 7,620 GPU node-hours at an average of 122.1 nodes on ROQUO. The abstract totals of 21,006 circuits, more than 239 hours of quantum processor time and 3.0 billion measurement outcomes cover all six calculations together; version 1 of the same preprint reports 9,200 circuits, more than 100 hours and 1.3 billion outcomes for its own smaller set and already reported binding energies. Cleveland Clinic's newsroom announcement of 9 September 2026 supplies the Gordon Bell finalist status and the announcement date; The Quantum Insider carried the same press release and is secondary to it. The team's earlier 303-atom protein simulation is arXiv:2512.17130. RIKEN's own June 2026 announcement supplies the identification of ROQUO, and the SC26 conference site supplies the dates of 15 to 20 November 2026 in Chicago. The judgments are this Monitor's own: the TRL 4 placement and the reasons for holding it below rung five, the decision to lead on the corrected binding-energy sign rather than the atom count, the reading of the abstract totals as a study aggregate rather than a repeat of version 1 at higher cost, the observation that accurate fragments with a wrong-signed assembled result locates the failure outside the quantum sampling, among the shared basis and embedding choices, without separating their contributions, the note that the mixed basis improves the assembled binding energy while its largest fragment agrees less closely with the classical reference, the caution that the 62.4-hour figure covers the bound-complex ground-state run alone and is not a whole binding energy, and the six procurement questions.