Simulating how a drug-target receptor cuts itself in two, with a quantum processor at the heart of the calculation
Earlier this year we previewed a biomolecular simulation pipeline coupling quantum computing with GPU-accelerated supercomputing. Today, QMatter and collaborators at the Technical University of Munich, University College London, Leibniz Supercomputing Centre, Ludwig-Maximilians-Universität München, NVIDIA, and IQM Quantum Computers release the full study:
The work applies the pipeline to a real open question in biology: how an adhesion G-protein-coupled receptor called ADGRG1 cuts its own backbone. The simulated system contains more than one million atoms, including the full-length receptor embedded in a realistic membrane, and the workflow brings in a 54-qubit IQM quantum processor and up to 1,200 NVIDIA H100 GPUs for the most chemically demanding part of the problem.
A receptor that cuts itself
GPCRs are one of the most important family of drug targets, with roughly a third of approved medicines acting on them. Adhesion GPCRs (aGPCRs) are the second-largest family within that group, yet no aGPCR has been targeted by an approved drug, despite links with several human diseases.
Most aGPCRs undergo autoproteolysis: the protein breaks one of its own peptide bonds, without help from any external enzyme, at a site inside its GAIN domain. ADGRG1 is a striking example. It has been implicated in platelet function and blood clotting, its mutations cause a rare disorder of brain development (bilateral frontoparietal polymicrogyria), and high ADGRG1 levels have recently been associated with aggressive triple-negative breast cancer. Understanding precisely how the cleavage happens could provide a basis for targeting this receptor family therapeutically. What makes the reaction hard to predict is that it is governed by energy barriers: how much energy the system must spend to from one chemical structure to the next. Those barriers depend sensitively on how electrons rearrange as bonds break and form, and on the surrounding protein, membrane and solvent. Previous simulations of aGPCR autoproteolysis used substantially simplified models: an isolated fragment of the rat receptor in water, without the membrane-spanning region. Cleavage is believed to take place in the membrane of the endoplasmic reticulum, so we chose to model it there within its full biochemical context.
Figure 1: ADGRG1 GPS autoproteolysis site
Interactive: the full-length, uncleaved ADGRG1 receptor embedded in an endoplasmic-reticulum-like membrane. Drag to rotate, scroll to zoom.
3Dmol.js failed to load for some reason. Please check your browser console for error messages.
Layer-by-layer breakdown of the system
A fully quantum-mechanical treatment of a million atoms is out of reach for any computer, classical or quantum. The pipeline instead applies different levels of accuracy to different regions of the system and runs each stage on the hardware best suited to it:
- The full system – 1,173,254 atoms. We built a model of the complete uncleaved human receptor, combining an AlphaFold prediction with an experimentally determined structure, embedded it in a membrane matching the composition of the endoplasmic reticulum, and solvated it. GPU-accelerated molecular dynamics showed the receptor remained stable and repeatedly adopted structures favourable for the reaction to start.
- The reactive region – 175 atoms. Around the cleavage site, we treated the catalytic residues and a corridor of ten water molecules with density functional theory, and reconstructed the step-by-step reaction pathway from the starting structure to the ester intermediate.
- The reactive centre – 52 qubits. At the bonds that actually break and form, we embedded a high-accuracy correlated wavefunction calculation inside the DFT region, itself polarised by the surrounding protein, membrane and solvent. This innermost problem was sized to fit the IQM Quantum processor: 26 electrons in 26 orbitals, mapped onto 52 of the 54 qubits available.
The pipeline is modular. Each stage exchanges well-defined data (geometries, Hamiltonians, quantum circuits and measurement data) so that better embedding schemes, solvers or quantum algorithms can be swapped in as they mature.
Quantum computers as accelerators
The quantum calculation uses quantum-selected configuration interaction (QSCI). The quantum processor prepares time-evolved quantum states and is measured to find the electron configurations that matter most. Those configurations are then collated and diagonalised on GPUs to compute the energy. For this work we developed an iterative version of QSCI. Rather than sampling the quantum computer once, information from each classical solution is fed back into the next round of quantum circuits, steering the sampling towards configurations the current solution is missing.
The workflow spans three systems. The classical steps were run on Wolpertinger, a high-performance computing (HPC) cluster that is part of BEAST, the Leibniz Supercomputing Centre’s (LRZ) testbed for future computing and HPC quantum integration (HPCQC). Wolpertinger is directly connected to Euro-Q-Exa, a 54-qubit IQM quantum computer at LRZ. The circuits were prepared on Wolpertinger and sent to the quantum processor, with around a million measurements per reaction structure, orchestrated through NVIDIA CUDA-Q – the open platform for hybrid quantum-classical supercomputing. During workflow execution, the compute-intensive GPU post-processing was run on 64 NVIDIA H100 GPUs of the BayernKI AI cluster (the Bavarian IT infrastructure for AI), under the same roof as the quantum computer at LRZ. To assess how this step scales we ran it separately on the NVIDIA Eos cluster: near-perfect parallel efficiency of 97.4% across 1,200 NVIDIA H100 GPUs.
What we found
- Electron correlation changes the answer. The predicted energetics depend strongly on how accurately electron correlation is described. One widely used density functional predicts a barrier of 36.5 kcal/mol for the first step of the reaction, while a mean-field calculation gives 14.0 kcal/mol, and the two methods disagree on the ordering of the intermediates. Because reaction rates depend exponentially on barrier heights, differences of this size decide whether a reaction happens on a biologically relevant timescale at all. The correlated treatment resolves the disagreement, recovering the ordering proposed in earlier work and lowering the overall barrier from 55.4 to 51.9 kcal/mol. The largest effect is in the first proton-transfer and bond-forming step, exactly where approximate methods are known to struggle.
- The quantum pipeline matches a state-of-the-art classical benchmark. For the three structures along the pathway, QSCI gives barriers of 25.9, 28.3 and 52.1 kcal/mol relative to the reactant, against 25.5, 27.5 and 51.9 kcal/mol from extrapolated heat-bath configuration interaction (HCI).
- The biology is consistent. Despite a substantially more realistic model, a different receptor and a different species, the results are broadly consistent with earlier simulations of the rat receptor ADGRL1, and support a picture of autoproteolysis as a one-time cleavage event. Agreement between independently constructed models strengthens confidence that the underlying chemistry is captured.
We are clear about what this does and does not show. This reaction is weakly correlated, so cutting-edge classical methods like HCI already perform exceptionally well, and we do not claim a quantum advantage here. That was not the aim. The aim was a general, transferable workflow applied to a biological problem of genuine interest. Notably, the quantum sampling recovered high-order electron configurations that become increasingly expensive for conventional methods to find. They carry little weight in this system, but they are exactly what matters in strongly correlated chemistry, where quantum sampling should help most.
What comes next
This study shows how quantum processors can sit inside large-scale scientific workflows as accelerators, alongside CPUs and GPUs, rather than as standalone machines. As quantum hardware and algorithms improve, the same pipeline can be pointed at more strongly correlated active sites, larger active spaces and new biomolecular targets, extending correlated electronic-structure methods to systems beyond the reach of conventional approaches alone.
Read the full preprint: https://arxiv.org/abs/2610.23108