International Conference on Circuits, Control, Communication and Computing (I4C)
Hybrid Quantum Graph Neural Networks for Molecular Toxicity Classification
Anurag Rai, G Nithish, Dasiga Venkata Ashish Kumar, Aishwarya JA
Abstract
Drug discovery depends heavily on predicting molecular toxicity early and classical ML still struggles here, because the interactions that make a molecule toxic are fundamentally quantum-mechanical. This paper builds a Hybrid Quantum Graph Neural Network (QGNN) that feeds GCN-learned molecular embeddings into a Variational Quantum Circuit (VQC). A 3-layer GCN produces graphlevel embeddings, which are angle-encoded into 4-8 qubit quantum states, and Pauli-Z expectations are concatenated with classical features before the final classifier. A quantum edge embedding is also introduced, encoding bondtype information directly into entangling gate parameters. On MoleculeNet's Tox21 (7,831 compounds, 12 endpoints) and ClinTox (1,478 compounds), 5-fold cross-validation shows the hybrid 4-qubit model with edge embedding achieves 0.762 ROC-AUC on Tox21 (+3.1% over 0.739 classical baseline), while the 8-qubit hybrid model reaches 0.741 on ClinTox (+6.3% over 0.697 baseline), both with batch size 64. Smaller batches (64) consistently outperform larger batches (256) at$3-17 \times$computational cost. Quantum edge embedding yields its largest improvement with only 15 additional parameters: +16.4% on ClinTox (batch 64). Quantum-only models achieve 83.7% parameter reduction while landing consistently just below the classical baseline. Across five folds, however, every one of these differences falls inside overlapping 95% confidence intervals$(p>0.05)$, so the gains reported here are consistent in direction but not individually significant at this sample size. Total overhead for hybrid models is 10-13% extra parameters with training time ranging from 1.2 to 54.6 seconds per epoch. External validation on the held-out BBBP benchmark (2,039 compounds), performed with the same architecture and no retuning, confirms that the gain transfers: the hybrid 4-qubit model reaches 0.702 ROC-AUC against 0.649 for the classical baseline (+8.2%), the largest relative gain across the three datasets. A mechanistic interpretability analysis attributes the edge-embedding effect to bond order reaching the model only through the entangling gates, where a learnable convex combination of perbond-type angles tunes entanglement continuously instead of pinning it at the CNOT maximum; the effect is positive on reactivity-driven toxicity endpoints and negative on passive permeability, matching the chemistry of each target.