ISSN print edition: 0366-6352 ISSN electronic edition: 1336-9075 Registr. No.: MK SR 9/7 Published monthly

Dipole moment prediction with graph neural networks and physics-informed models: a comparative evaluation on QM9 and traditional Chinese medicine molecules

Xiong Li, Guoqiang Bian, Tao Yang, Kongfa Hu, Renli Xu, Zuojian Zhou, and Chenjun Hu

School of Artificial Intelligence and Information Technology, Nanjing University of Chinese Medicine, Nanjing, China

E-mail: renlixu@njucm.edu.cn

Received: 30 October 2025  Accepted: 5 January 2026

Abstract:

This study systematically compares six deep learning architectures—three graph neural networks (MPNN, GAT, GIN) and three physics-informed variants (PhysNet, PhysNet-ens5, PhysNet-Lite)—for molecular dipole moment prediction. Models are first evaluated on the QM9 benchmark, then assessed for generalization to a curated set of 50 Traditional Chinese Medicine molecules (27 overlapping with QM9, 23 structurally novel). While physics-informed models achieved the highest accuracy on QM9, GAT demonstrated the most robust transfer to unseen TCM structures, with the lowest performance degradation (gap ratio: 1.48 ×). Physics-informed models showed potential but exhibited varying generalization behavior, indicating that architectural design and hyperparameter optimization significantly influence out-of-distribution performance. This comparative analysis provides practical guidance for model selection in TCM property prediction and establishes a framework for evaluating generalization in molecular machine learning.

Keywords: GNN; Physics-informed models; Dipole moment; QM9; Traditional Chinese medicine

Full paper is available at www.springerlink.com.

DOI: 10.1007/s11696-026-04657-0

Chemical Papers 80 (4) 3885–3894 (2026)