Researchers from Singapore's Agency for Science, Technology and Research and the National University of Singapore ran a hybrid molecular-docking method on IBM quantum hardware. Their August 24 preprint uses full-basis encoding to fit more optimization variables into fewer qubits.
The experiment recovered the same selected contacts as classical calculations for two small protein-ligand problems. This is a hardware feasibility result, not a drug-discovery breakthrough or a quantum speedup.
Turning docking into a graph problem
Molecular docking searches for plausible positions where a ligand binds to a target protein. The search grows difficult when a ligand has many possible interaction points and orientations.
The team converts each possible pharmacophore contact into a graph vertex. Compatible contacts connect through graph edges. A valid docking pose becomes a clique, and the preferred pose becomes a maximum vertex-weighted clique.
This mapping turns part of docking into a discrete optimization problem. Classical solvers already handle such problems. The quantum question is whether a compact variational encoding helps solve useful instances under near-term hardware limits.
One qubit holds three variables
Standard variational mappings assign one binary variable to one qubit. Full-basis encoding uses the three single-qubit Pauli expectation values, X, Y, and Z, as separate relaxed variables. A problem with N variables therefore needs roughly ceil(N/3) qubits before other hardware constraints enter.
The method applies a hyperbolic tangent to measured expectation values, then optimizes the resulting cost function on a classical computer. A warm-start stage uses a stochastic imaginary-time-inspired procedure before gradient descent. The quantum processor evaluates the circuit while the classical optimizer updates its parameters.
The paper also proves a useful property of this encoding: at least one global minimizer has a pure product-state representation. Entanglement is not required for the mathematical optimum under this formulation. That does not remove hardware noise or classical optimization cost, but it explains why shallow circuits fit the method.
What IBM hardware recovered
The researchers built two instances from experimentally resolved protein structures:
- PDB structure 1stp, with 18 graph variables and a six-qubit full-basis circuit.
- PDB structure 9aw2, with 14 graph variables and a five-qubit full-basis circuit.
They ran the circuits on ibm_kingston, a 156-qubit Heron R2 processor. The experiment used three variational layers and standard readout mitigation. Hardware measurements recovered the same vertex selections as exact classical simulations and matched the known clique structure for both examples.
The qubit compression is the main engineering result. Full-basis encoding put 18 variables into six qubits and 14 variables into five. A one-variable-per-qubit mapping would need 18 and 14 qubits before accounting for connectivity and routing.
Why this matters for quantum chemistry
Near-term processors struggle with deep circuits, noise, and limited qubit counts. A method that reduces qubit demand gives researchers more room to test structured optimization problems on current devices.
The docking instances also use real protein-ligand structures rather than an abstract graph. That makes the demonstration closer to a scientific workflow than a generic Max-Cut test. The quantum portion still handles a small encoded subproblem, while classical code constructs the graph, performs warm-start work, updates parameters, and interprets the result.
This fits the pattern in our protein simulation coverage. Current quantum chemistry demonstrations often combine a small quantum calculation with a much larger classical pipeline. The useful question is which part moves to hardware and whether that part improves the complete workflow.
What the preprint does not prove
The study does not report a quantum advantage. Both instances are small, and classical simulation knows the target solutions. Matching those solutions shows successful recovery under the chosen conditions. It does not show lower runtime, lower energy use, better docking accuracy, or performance on large flexible molecules.
The preprint also leaves scale as an open question. Full-basis encoding reduces qubit count by sharing three observables across one qubit, but those observables remain coupled by the Bloch-sphere constraint. Larger graphs bring harder optimization, more measurements, and more opportunities for noise to alter the recovered clique.
The next useful test is a larger blind benchmark against strong classical docking software. Until then, this result belongs in the near-term quantum optimization category: a careful small experiment showing how hardware and algorithm design meet in the middle.