arXiv:2610.04588
Variational Quantum Attention for Molecular Graph Learning
Yu-Cheng Lin, Yu-Chao Hsu, Tai-Yue Li, Nan-Yow Chen, Samuel Yen-Chi Chen
Abstract
Molecular property prediction is central to computational drug discovery, where graph neural networks learn to weight neighboring atomic environments during message passing. Yet it remains unclear how variational quantum circuits alter learned attention behavior in molecular graphs. We introduce an edge-aware variational quantum attention mechanism for molecular graph learning, in which the receiving atom, neighboring atom, and connecting bond jointly determine the quantum attention state. Across five molecular property and bioactivity prediction tasks, QGAT achieves competitive performance relative to GATv2, with a consistent improvement on BBBP across all six evaluated circuit ansatzes. We further compare how the quantum and classical attention scores weight molecular structure beyond accuracy. In the Verubecestat BACE1 inhibitor series, QGAT achieves a higher Spearman correlation than GATv2 and assigns positive attributions to several structural changes consistent with reported structure-activity relationships (SARs). This case study shows that the two attention mechanisms can exhibit different prediction and attribution behavior across structurally related BACE1 analogues, while broader validation is required to determine how consistently these differences generalize across chemical series and targets. Circuit ablations further show that performance depends on the circuit design. Together, these results show that variational quantum attention can serve as a viable alternative molecular attention parameterization while inducing circuit- and chemistry-dependent behavior distinct from a matched classical scorer.
AI Summary
The paper asks whether variational quantum circuits can replace classical attention scorers in molecular graph networks while changing predictive and molecular-attribution behavior. Its edge-aware quantum graph attention network (QGAT) combines projected receiving-atom, neighboring-atom, and bond features, amplitude-encodes the resulting 64- or 128-dimensional vector into 6 or 7 qubits, and uses scaled Pauli-$Z$ expectation values as attention logits. Message passing, aggregation, and prediction remain classical. Evaluation covers five molecular classification and regression datasets using a fixed 10% scaffold-held-out test set and five training/validation scaffold folds. Circuits are trained using exact state-vector simulation. The strongest evaluated QGAT configuration per task is competitive with a matched GATv2 baseline: BBBP ROC-AUC increases from $0.6808\pm0.0111$ to $0.7157\pm0.0095$, with all six tested ansatzes exceeding the baseline mean. Other differences are smaller or mixed; FreeSolv slightly favors GATv2. QGAT uses approximately 0.5–1.3% fewer trainable parameters. Circuit ablations show task-dependent performance, with optimal depths of one to three blocks rather than monotonic improvement with depth. Substructure masking, attention visualization, and integrated gradients characterize differences in molecular weighting. On the external Verubecestat BACE1 inhibitor series, QGAT achieves Spearman correlation 0.8264 versus 0.7954, but worse RMSE. Several attribution directions agree with reported structure–activity relationships, while both models miss others. Stated limitations include ligand-only inputs, no hardware evaluation, and primarily qualitative chemical interpretation centered on one series; the results do not establish generally superior interpretability.
Error found: The headline prediction table reports the strongest observed QGAT configuration for each task without documenting independent validation-based configuration selection, leaving its best-configuration gains vulnerable to selection bias.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.