Quantum-derived molecule scores change a drug-screening shortlist—not prove an advantage

Same molecules, different descriptions

Alejandro Giraldo and colleagues studied 354 compounds with existing active-or-inactive labels for DYRK1A, an enzyme investigated as a drug target. Their experiment was retrospective: it asked models to predict labels already available to the researchers, rather than discovering and experimentally confirming new inhibitors.

The comparison began with two ways of describing each molecule. The conventional representation contained 1,080 numerical molecular descriptors. The alternative came from encoding those descriptors into a quantum circuit running on 71 qubits on IBM’s ibm_boston processor.

Measurements in three directions for each qubit produced 213 values per molecule. These became inputs to an ordinary classical machine-learning algorithm, a radial-basis-function support vector machine. The quantum processor supplied a different representation; it did not make the final classification itself.

The researchers also simulated the circuit using a matrix product state method, which can represent suitably limited entanglement without storing every amplitude of the full quantum state.

What changed on the shortlist

For the main comparison, the team trained on 264 molecules and tested on 90. Hyperparameters were selected through cross-validation within the training set. Both models used their default decision threshold to mark compounds active or inactive.

The conventional model ranked compounds better overall. Its ROC-AUC was 0.8790, compared with 0.8255 for the quantum-derived representation. This measure describes how well scores place active compounds ahead of inactive ones across possible thresholds; it does not specify which molecules enter a particular shortlist.

At the chosen threshold, the lists differed on 15 compounds. The quantum-derived model alone selected six: two labeled active and four inactive. The conventional model alone selected nine: six active and three inactive.

That makes the practical trade-off unusually clear. Adding the quantum-only selections to the conventional shortlist would recover two additional labeled actives while requiring tests of four additional inactive compounds. Replacing the conventional list with the quantum-derived list would instead lose six actives that the conventional model selected.

These are consequences calculated from existing labels—not results from a new laboratory screening campaign.

Known chemistry, not new medicines

The two active compounds selected only by the quantum-derived model were 8d, a previously reported DYRK1A analogue, and XMD14-124, a research probe associated with a published protein structure. Neither is an approved drug.

Both had training-set neighbors within the study’s defined applicability domains. Their disagreement therefore was not simply a case of one model encountering molecules far outside its familiar territory.

There was an important difference between them, however. XMD14-124 sat very close to the quantum-derived support vector machine’s decision threshold, making its binary classification particularly sensitive to small score changes. The paper also reports differing published potency measurements for 8d, emphasizing that a reused binary activity label does not resolve differences between assay protocols.

Repetition supports disagreement, not superiority

Across 100 additional train–test partitions of the same dataset, 8d was selected only by the quantum-derived model in 19 of its 21 test appearances. XMD14-124 showed that pattern in 20 of 27 appearances. The quantum-derived model had the higher ROC-AUC in 21 of the 100 partitions.

These repetitions suggest that the shortlist differences were not confined to one split. They are not independent experimental confirmations: the researchers reused the same compounds and already-measured quantum vectors.

Combining the representations gave an exploratory ROC-AUC of 0.9022, slightly above the best conventional-only result across the classifier families tested. But classifier-family selection for those analyses used test-set performance, limiting what that improvement establishes. Neither combination retained both highlighted quantum-only actives.

What remains unresolved

On the main 90-compound test set, the ranking difference did not reach conventional statistical significance: the reported DeLong test gave p = 0.117. That does not establish equivalent performance.

The authors also investigated hardware noise, but its causal role remains uncertain. Measurement-shot counts, simulation bond dimension and mitigation details are absent from the reported record, limiting reproducibility and interpretation.

The November 10 preprint does not establish quantum advantage, clinical benefit or a validated screening improvement. Its narrower finding is useful: changing a molecular representation can change which known actives reach a proposed assay list—even when the overall ranking score looks worse.