Year: 2026
Type: preprint
Abstract: Parameterized quantum circuits (PQCs) are widely proposed as feature transformations in quantum machine learning (QML) pipelines. However, few molecular-QML studies combine strong classical baselines, repeated training runs, and controls that separate a quantum branch from its trainable anglegenerating encoder. We benchmark three-qubit PQCs combined with two pretrained molecular embeddings, MiniMol and CheMeleon, for lipophilicity and solubility prediction. Five model variants, including a classical baseline, two hybrid quantum-classical architectures, a quantum-only model, and a frozen-branch control, were evaluated on two datasets, two embeddings, and five random seeds, giving 100 runs. No tested quantum variant showed a statistically detectable improvement over the classical baseline after falsediscovery-rate correction across all 16 paired comparisons. The frozen-branch control also showed no detectable advantage over the classical baseline. This control freezes the full angle-generating quantum branch, not an independent set of variational circuit weights. A separate one-qubit depolarizing-noise sweep showed no detectable change in accuracy over the tested error range. Quantum variants required 18–37 times longer training time than the classical baseline on the same platform. These results set a benchmark for small parameterized QML layers applied to the pretrained molecular representations and ADME tasks examined here.
Source: ChemRxiv
Institutions: Northeastern University, Northwestern University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204834928