A Dimension-Matched Quantum Feature Map for Hierarchical Gene Ontology Term Prediction: Attributing the Gain on CAFA6

<title>Abstract</title> <p> <bold>Background and motivation.</bold> Gene Ontology (GO) term prediction assigns hierarchical functional labels to protein sequences, and fewer than one percent of known sequences carries experimental annotation. <bold>Research gap.</bold> Researchers increasingly apply quantum feature maps to biological classification, but the reported comparisons leave one confound uncontrolled: a circuit mapping inputs to a longer vector of expectation values is compared against a classical model on the original dimensions, so any gain conflates the quantum map with a wider representation and a bounded non-linearity. <bold>Proposed method.</bold> We present QEPP, a hybrid quantum–classical feature extractor in which a shallow angle-encoded circuit with a correlation-derived entangling topology feeds a single shared multi-label head, with ontology consistency enforced in the loss and at inference. <bold>Experimental evaluation.</bold> We evaluate the proposed framework on a 180-term CAFA6 subset of 40,000 proteins under a cluster-disjoint split, against three baselines including a frozen random cosine map of identical output width trained with the identical head, split, label set and threshold protocol, reporting, AUPR and over five seeds. <bold>Main results.</bold> Our method improves on the width-matched control on AUPR (+ 2.24%, ) and (+ 1.02%, ) in five of five seeds, at 49 times the control’s wall-clock; representational width alone accounts for 85% of the total gain. Two components return null results: the structured entangler is indistinguishable from a linear ladder, and variational tuning is inert. <bold>Contribution.</bold> The presented approach is the first quantum feature map applied to GO annotation and the first dimension-matched evaluation of one on a biological prediction task. </p>