Feasibility and optimum recovery in warm-start quantum optimization for a drug-response model on a trapped-ion processor

AI summary

Overview

Research area: Quantum optimization (warm-start QAOA) benchmarked against classical solvers, applied to a drug-combination dose-selection model derived from the AstraZeneca–Sanger Drug Combination DREAM dataset and executed on a trapped-ion processor.

Technical level: Intermediate. The paper assumes familiarity with QAOA, QUBO-style cost functions and sampling baselines, but its central argument — that feasibility gains are not optimization gains — is stated in plain terms and backed by explicit controls.

Scope (one sentence): The authors test whether warm-start QAOA actually recovers optimal dose assignments on trapped-ion hardware, or merely produces more valid assignments than random bitstrings, using exact enumeration of feasible sets up to 4,900 assignments and 6–35 qubits.

What This Paper Is About

Many quantum optimization papers claim success by showing that a circuit outputs valid (constraint-satisfying) solutions more often than random bitstrings. For constrained problems this is a weak test, because a sampler that only respects constraints can look impressive without ever finding good solutions. The authors take archived warm-start QAOA circuits for a seven-compound cancer drug-response model, run on IonQ Forte-1, and ask a stricter question: compared with simply sampling uniformly from the feasible set, does the quantum circuit actually put more probability on the optimal dose assignment — and does hardware preserve that advantage?

Key Contributions

1. A three-way control design for warm-start QAOA. Each execution is compared against its own product-state initialization (warm-start-only), against uniform sampling over all 2^q bitstrings, and against uniform sampling over the feasible set only. Optimum probability is also converted into a probability of finding the optimum at least once in 200 draws, so hardware and analytic baselines are compared on the same footing.

2. Exact reference optima by full enumeration. The authors enumerate feasible assignments directly (using the closed form |F| = C(n,2)D² + C(n,3)D³ for instances where every compound has a distinct mechanism class), establishing exact global optima for every instance, including a 35-qubit encoding with 4,900 feasible assignments. Noiseless statevector references are computed for six original circuits and six wider circuits at up to 35 qubits.

3. An audit of 25 archived hardware tasks. 18 distinct circuits across 14 optimization instances, totaling 5,300 shots (24 executions of 200 shots and one of 500), are analyzed against hardware, noiseless and classical baselines, including repeated executions of identical circuits.

4. Classical constructive baselines and a feasibility-preserving mixer. Greedy addition plus feasible local improvement and simulated annealing, both capped at 200 unique objective evaluations, are run on every model; a one-layer Grover-mixer QAOA is simulated on ten models to test whether constraint-preserving mixing changes the picture.

Main Findings

• Feasibility enrichment is real but mostly an encoding effect. Hardware returned valid assignments in 29.5%–88.0% of shots, versus 1.10%–31.25% for uniform random bitstrings, in every one of the six reference instances. At 15 qubits only 1.10% of bitstrings are feasible, so even a 59.2× enrichment left over a third of hardware shots invalid.

• Ideal QAOA beats its own initialization far more often than it beats feasible sampling. Ideal QAOA exceeded its product prior in 11 of 12 reference circuits, but exceeded uniform feasible sampling in only 4 of 12.

• Optimum recovery on hardware is sparse. Hardware found the unique optimum (the same three compounds at their highest encoded dose, score 1.769364) in only 2 of the 6 reference batches — 5 times each at 6 and 8 qubits, and never in the other four. A zero count yields a Wilson upper endpoint of about 1.88% per shot, not a proof of zero probability.

• High feasibility coexists with zero optimum recovery. At 15 qubits, hardware returned 130 valid outputs and no optimum. Ideal optimum probability there was 0.247%, versus 0.144% for the prior and 0.278% for uniform feasible sampling, giving 200-draw success probabilities of 39.0%, 25.0% and 42.7% respectively; uniform bitstrings gave 0.61%.

• Ideal predictions diverge sharply from hardware. The 10-qubit ideal circuit predicts a 94.0% chance of an optimum within 200 draws, yet hardware returned none in that batch.

• Hardware sometimes falls below its own noiseless circuit. Hardware lost feasible probability relative to the corresponding ideal circuit at 18, 21 and 25 qubits, with every pointwise hardware interval at these settings excluding the ideal value; both 20-qubit intervals included it. Ideal feasibility was 73.16% at 18 qubits and 73.44% at 21 qubits.

• At the widest circuits, both circuit and execution limitations appear together. At 30 qubits (2,875 feasible assignments) ideal feasibility was 33.39% against a prior of 33.64%, and hardware returned 13 of 200 valid outputs. At 35 qubits (4,900 feasible assignments) ideal feasibility was 35.85% against a prior of 35.75%, and hardware returned 7 of 200 (Wilson interval 1.7–7.0%), with a best normalized gap of 40.4%. Ideal optimum probability there was 0.01197%, below the prior's 0.01217%.

• Classical search solves everything the hardware was pointed at. Greedy search solved all six reference instances in 21–81 unique objective evaluations and all 14 hardware-associated models in 21–200, versus 20–4,900 for exact enumeration. Annealing reached the optimum in 15–100% of 100 seeds across the 14 hardware models (98–100% on the original six, median first hit 11–87.5 evaluations); at 35 qubits it succeeded in 15 of 100 seeds, and greedy reached the optimum at its 200-evaluation cap.

• Panel simulations show the sampling gain is model-dependent. Ideal QAOA beat both sampling controls in all 12 runs of a four-panel simulation (each panel 162 feasible assignments; optimum probability 2.23–3.43% versus 1.26–1.48% for initialization and 0.617% for uniform feasible sampling). Extending to wider panels, ideal QAOA beat uniform feasible sampling in all three seeds for 28 of 35 CAMA-1 panels (80%) and in none of four 647-V panels. Greedy solved all 35 CAMA-1 panels and 3 of 4 647-V panels (75%); annealing solved every panel in all 100 seeds.

• Repeated executions of identical circuits vary. Feasibility ranged from 35.0% to 44.0% at 21 qubits and 42.0% to 49.5% at 18 qubits; the two 6-qubit runs were closer at 88.0%–88.4% with 200 and 500 shots.

• A Grover mixer guarantees feasibility and improves on feasible sampling, but not uniformly on penalty QAOA. Across the six original instances and four original panels, its optimum probability ranged from 1.17% to 27.24%, exceeding archived penalty-QAOA on all six reference instances but falling below trained penalty-QAOA on all four panels.

• Score-weight sensitivity leaves the headline comparison intact. Across λ swept from 0.3 to 0.9 in steps of 0.15, all 14 hardware models retained the same optimum; greedy recovered it in 61 of 70 model/weight combinations; retrained QAOA beat feasible sampling in 52 of 60 runs and beat its product prior in 59 of 60.

Methodology in Plain English

The authors built an optimization problem out of real laboratory measurements: how sensitive the CAMA-1 cancer cell line is to seven compounds individually, and how strongly each of the 21 pairs of compounds interacts. A solution is a choice of two or three compounds with one dose level each; invalid choices (no compound, two doses of the same compound) are penalized in the cost function. The score combines a fitted single-agent inhibition estimate, a class-weighted dose proxy, and pairwise synergy scaled by the product of dose fractions.

Because the feasible sets were small enough (20–4,900 assignments), the team could list every valid assignment and score it exactly, which gave them ground-truth optima rather than estimates. They then compared four things on identical instances: the quantum hardware output, a noiseless simulation of the exact same submitted circuit, a classical sampler that only ever draws feasible assignments, and classical searches (greedy local improvement and simulated annealing) with a strict 200-evaluation budget. They also reconstructed the initial product state to check how much of the quantum result was already present before any cost evolution, and simulated a Grover-mixer variant that cannot leave the feasible set by construction.

For the wider circuits they had to compute full statevectors at 30 and 35 qubits, done on a Pawsey node using single-precision amplitudes in Gray-code order, validated to within 10⁻⁷ against double precision at 25 qubits.

Why This Matters

The paper supplies a concrete counterexample to a common evaluation practice: reporting that a quantum optimizer "satisfies constraints more often than random" while the appropriate baseline — sampling valid solutions uniformly — does better. It also documents an execution shortfall in which hardware underperforms its own noiseless circuit, which matters for anyone interpreting archived hardware results.

Real-world applications:

• Drug-combination screening: prioritizing which compound pairs and dose levels to test next, where each experiment is expensive and the candidate space is combinatorial.
• Constrained portfolio and resource-allocation problems: the warm-start and constraint-preserving mixer techniques studied here originate in this setting.
• Benchmarking quantum cloud services: the three-control framework (prior, uniform, uniform-feasible) is directly reusable for auditing vendor hardware claims.
• Mixed-integer experimental design: any setting with structural constraints (pick k of n items, one level each) and a cheap-to-enumerate feasible set.

Industry relevance: the results give quantum hardware providers and cloud platforms a more honest reporting template, and give potential adopters a clear signal that for problems of 20–4,900 feasible assignments, well-tuned classical search remains a strong baseline that quantum circuits did not beat here. The authors explicitly support no computational-advantage claim.

Future Directions

• Isolate the source of the execution shortfall. The authors state that preparation controls, calibration records and native compiled circuits would be needed to identify why hardware feasibility dropped below the ideal circuit at 18, 21 and 25 qubits; native compiled CNOT counts were unavailable, so only submitted counts (30–1,190) could be reported.
• Account for the hardware cost of preparing the feasible state. The Grover-mixer results come from direct state construction in simulation; the authors note its hardware preparation cost remains to be assessed before the comparison is meaningful on a device.
• Separate mixer effects from training effects. Because initialization and training also change between penalty QAOA and Grover-mixer QAOA, the current comparison does not isolate the mixer alone.
• Test whether the CAMA-1 pattern transfers. The 647-V panels showed no sampling gain in any of four panels, raising the open question of which model features predict when warm-start QAOA beats feasible sampling.

Target Audience

Quantum computing researchers benchmarking near-term optimization algorithms, especially those working on constrained or warm-start QAOA and on trapped-ion hardware. It is equally useful for computational biologists and cheminformatics researchers evaluating whether quantum methods add value over enumeration and local search on drug-combination models, and for reviewers and practitioners who need a template for fair classical baselines. Readers should be comfortable with QAOA terminology; the statistical and biological framing is accessible to a general technical audience.

Authors’ abstract

On a seven-compound drug-response model, warm-start quantum approximate optimization (QAOA) on IonQ Forte-1 returned valid assignments more often than random bitstrings, but this alone did not show effective optimization. Ideal QAOA raised optimum probability above uniform feasible sampling in only four of twelve reference circuits. Hardware often fell below its own noiseless circuits, while greedy search solved all hardware models within 200 objective evaluations. Expanded simulations showed a gain over feasible sampling in 28 of 35 CAMA-1 panels and none of four 647-V panels. Annealing solved all these panels in every seed. A Grover mixer preserved feasibility and improved optimum probability over feasible sampling in all ten tested models. We analyzed 25 completed tasks containing 5,300 shots from 18 circuits and 14 instances. The encodings use 6-35 qubits and at most 4,900 feasible assignments, which we enumerated to establish exact optima. Noiseless references now cover both the original six circuits and six wider circuits. At 35 qubits, with 4,900 feasible assignments, ideal feasibility was 35.85%, compared with 7 of 200 valid hardware outputs. Its ideal optimum probability was below both sampling controls. The circuits sample assignments in a model built from measured single-agent and pairwise responses.