Quantum-Embedded Graph Neural Network Architecture for Molecular Property Prediction

Accurate
prediction of molecular properties is crucial for accelerating
the development of new drugs, and quantum machine learning (QML) holds
great promise in this domain. A typical QML pipeline comprises two
core stages: encoding classical data into quantum representations
followed by training and prediction using quantum computing-based
machine learning (ML) models. In this article, we focus on the initial
encoding stage and propose an effective quantum feature extraction
approach for molecular graph data, introducing quantum node embedding
and quantum edge embedding methods. We developed a hybrid quantum–classical
ML framework and implemented several quantum-embedded graph neural
network (QEGNN) models to evaluate the proposed method. Experiments
conducted on three benchmark data sets with diverse molecular property
prediction tasks demonstrate that QEGNN models consistently achieve
higher accuracy, improved stability, and significantly reduced parameter
complexityhallmarks of quantum advantage. Furthermore, we
validate the reliability of the quantum embedding approach on the
superconducting quantum processor “Wukong,” showing
that the models retain stable performance even under the constraints
of current noisy quantum hardware. This work highlights the potential
of QML and paves the way for the development of universal QML models.