Fragment-Enhanced Graph Neural Networks for Target-Specific Molecular Generation and Property Prediction
Xu Gao, C. Xian F. Zhang, Mengfan Yan, Xianyu Zuo, Kecheng Yang · 2025
Molecular generation and property prediction are core tasks in drug discovery. However, existing methods heavily rely on labeled data and often fail to capture the multi-level semantic information of molecular graphs. To address these limitations, this paper proposes a fragment-enhanced Graph Neural Network (GNN) combined with a Conditional Variational Autoencoder (CVAE) for molecular generation. First, molecular fragments (motifs) with chemical significance are extracted using the retrosynthesis-based BRICS algorithm and additional rules. A two-stage framework is then designed: in the first stage, a fragment-enhanced GNN is pre-trained on unlabeled molecular data through self-supervised learning; in the second stage, the CVAE model incorporates protein pocket information as a condition to generate target-specific molecules. Experimental results demonstrate that the proposed method significantly outperforms traditional GNN-based methods in molecular property prediction tasks. Moreover, the method achieves strong overall performance in molecular generation, including an average synthetic accessibility (63.7%), average compliance with Lipinski's Rule of Five (4.991), and a PB-Valid success rate of 42%, validating the chemical validity and target specificity of the generated molecules. This approach provides a novel and effective solution for molecular generation and representation in drug discovery.