Matched-Control Benchmarking of Quantum Machine Learning Models for Hierarchical Protein Function Prediction

Predicting protein function requires assigning hierarchical Gene Ontology (GO) terms to sequences with limited experimental annotation. We tested whether quantum machine learning improves this task when each model is compared with a classical control matched to its architecture or input representation. Five quantum models spanning variational circuits, quantum features, and fidelity kernels were evaluated on 40,000 proteins from the CAFA6 training release, using frozen 480-dimensional ESM-2 embeddings, 180 GO terms, and an internal cluster-disjoint train–validation–test split. Kernel bandwidths were selected through a training-only concentration screen (Gate-0), which also estimated shots required per kernel entry. Across five seeds, the standalone quantum models scored below their matched controls by 0.0023–0.0847 in aspect-mean maximum F-measure; the 48-layer re-uploading kernel was admitted in only two seeds and lost in both. Freezing the variational circuit did not reduce performance. Late fusion of quantum and classical predictions improved on cosine nearest neighbours by 0.0034, but kernel-level mixing selected a purely classical weight, and the fused score (0.6697) remained below that of a classical RBF nearest-neighbour model (0.6719). Finite-shot simulations preserved this ordering, although the quantum kernel ridge model was sensitive to shot noise. These results provide no evidence of a standalone quantum advantage on this benchmark. The fusion gain is consistent with complementary prediction errors, but its mechanism remains to be tested against a matched classical ensemble.