arXiv:2610.23108
Multiscale Quantum-Classical Simulation of Protein Autoproteolysis in a Million-Atom Membrane Environment
Alexander Orlov, Alexis Ralli, Tim Weaving, Michael Williams de la Bastida, Reuben Martin, Zohim Chandani, Minh Chung, P. V. Sriluckshmy, Fedor Simkovic, Luigi Iapichino, Peter J. Love, Christian B. Mendl, Andrea Townsend-Nicholson, Peter V. Coveney
Abstract
Predicting chemical reactions in biomolecules is challenging because reactive centers are embedded within complex environments that strongly influence their energetics and mechanisms. Here, we present a hybrid quantum-classical workflow that combines molecular dynamics, density functional theory, and correlated wavefunction methods, tightly integrating CPUs, GPUs, and quantum processing units (QPUs) to enable simulations across multiple length and accuracy scales. We apply this approach to ADGRG1 autoproteolysis, a maturation reaction in an adhesion G protein-coupled receptor that has not previously been studied using correlated electronic-structure methods. Our simulations treat a biologically realistic system containing more than one million atoms, including the full-length uncleaved receptor embedded in an endoplasmic-reticulum membrane, while incorporating a 54-qubit IQM quantum processor and up to 1,200 NVIDIA H100 GPUs. These results demonstrate how quantum computers can serve as accelerators within scalable scientific workflows, extending correlated electronic-structure methods to biologically relevant systems.
AI Summary
The paper examines whether a multiscale quantum–classical workflow can evaluate correlated electronic energies for ADGRG1 autoproteolysis in a full-length, membrane-embedded receptor. The 1,173,254-atom environment is treated classically; reaction-path reconstruction uses a 163-atom QM region, and projection-based embedding reduces the correlated calculation to CAS(26,26), encoded on 52 qubits of a 54-qubit processor. A proposed N–O acyl-shift pathway is reconstructed through constrained QM/MM scans. Four representative structures are evaluated using embedded DFT, Hartree–Fock, classical heat-bath configuration interaction (HCI), and iterative quantum-selected configuration interaction (QSCI). QSCI feeds classical subspace solutions back into stochastic time-evolution circuit compilation and combines quantum samples with occupancy-biased classical configuration expansion. Each structure uses approximately $10^6$ measurement shots; classical diagonalization is distributed across up to 1,200 H100 GPUs. Relative to the reactant, QSCI gives electronic energies of 25.9, 28.3, and 52.1 kcal/mol for the three subsequent structures, versus extrapolated HCI values of 25.5, 27.5, and 51.9 kcal/mol. Correlation restores the intermediate ordering obtained with Hartree–Fock but reversed by the embedded DFT calculation, and reduces the overall electronic energy rise from 55.4 to 51.9 kcal/mol. These are single-point electronic energies on DFT-derived geometries, not activation free energies or fully relaxed correlated profiles. Two intermediates require constraints, and reoptimizations retain residual forces. The wavefunctions are predominantly single-reference, with 92–94% reference weight; HCI performs comparably, so no quantum advantage is demonstrated. At larger configuration spaces, quantum data contribute mainly through orbital occupancies rather than directly sampled determinants.
Central claim may not hold: The abstract presents quantum processors as demonstrated accelerators, whereas the reported results establish their integration into a multiscale workflow without demonstrating a runtime or resource benefit over classical computation. The QPU-assisted solver does not demonstrate improved configuration-selection efficiency over classical HCI: larger-space energies are slightly worse, small-space barrier improvements arise from error cancellation, and recovered high-order configurations contribute negligibly to the energy.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.