QBioFusion-QSAR: Morgan-Anchored Quantum Multiple Kernel Learning for Small-Data Ligand Classification

QBioFusion-QSAR introduces a quantum multiple kernel learning framework combining quantum fidelity kernels with traditional Morgan fingerprints for drug discovery classification tasks. On a 54-molecule benchmark, the hybrid approach modestly improved accuracy and correlation metrics, though statistical validation across multiple random partitions showed gains were not consistently significant beyond classical methods.

QBioFusion-QSAR addresses a persistent challenge in computational chemistry: improving predictions when training datasets are small and molecular analogues exhibit conflicting activity labels. The research explores whether quantum computing kernels can augment classical chemoinformatics models, specifically testing this hypothesis on the PsychLight-A benchmark containing 54 molecules with psychoactive properties.

The framework combines Morgan/Tanimoto fingerprints—a well-established molecular similarity metric in drug discovery—with quantum fidelity kernels constructed from RDKit, Mordred, and Deep-PK descriptors. Initial results appeared promising, with accuracy improving from 81.5% to 83.3% and Matthews correlation coefficient rising from 0.613 to 0.645. Notably, the quantum enhancement correctly reclassified two nitrogen-methylated compounds that had been misclassified, suggesting quantum kernels capture activity-cliff phenomena where small structural changes cause large activity shifts.

However, rigorous statistical validation revealed critical limitations. When researchers repeated the five-fold cross-validation across ten random data partitions, the quantum multiple kernel learning approach failed to consistently outperform the simpler Morgan/Tanimoto baseline on mean Matthews correlation coefficient. Bootstrap confidence intervals for the improvement spanned zero, indicating the gains were not statistically significant and potentially attributable to chance rather than genuine quantum advantage.

This research demonstrates both the promise and peril of quantum machine learning in chemistry. While quantum kernels theoretically encode structural information differently than classical metrics, small datasets make it difficult to distinguish genuine improvements from statistical noise. The work's transparent audit trail—identifying which specific molecules benefited from quantum enhancement—sets a methodological precedent for future quantum chemistry applications, emphasizing validation rigor over optimistic preliminary results.

• →Quantum kernel learning improved accuracy on a 54-molecule drug discovery benchmark but gains were not statistically significant across multiple validation runs.
• →The framework successfully identified activity-cliff molecules where quantum similarity metrics outperformed classical Morgan fingerprints.
• →Small-data QSAR studies remain challenging; quantum computing advantages are difficult to validate without sufficient statistical power.
• →QBioFusion-QSAR provides an auditable methodology for isolating quantum-kernel contributions in molecular classification tasks.
• →Results suggest practical quantum advantage in chemoinformatics requires either larger datasets or more specialized problem domains.