MSF-HierGNN: a multi-source substructure-fusion hierarchical GNN method and web server to predict molecular property for drug design

Accurate prediction of the biological and physicochemical properties of molecules is of great significance in shortening the process and decreasing the failure rate of drug design. Thus, previous studies have established several benchmark datasets and developed several Graph Neural Networks (GNNs) based artificial intelligence (AI) predictive methods. However, since these methods encounter challenges such as incomplete representation of molecular hierarchical structures, insufficient exploration of local topological features, and under-extraction of correlations among atomic features, our study proposes a graph neural network that combines hierarchical pooling with localized substructure modeling named MSF-HierGNN (Multi-Source Substructure-Fusion Hierarchical Graph Neural Network) to solve these problems. Firstly, MSF-HierGNN constructs a more comprehensive and multiscale graph-level representation by preserving chemically salient structures and fusing multiple sources of substructure. Secondly, the model can more comprehensively represent molecular substructure features by integrating multiple fragmentation algorithms and incorporating five molecular fingerprints. Additionally, the model further captures correlations among atomic features by increasing the message-passing mechanism. Finally, we validate the model's effectiveness using public benchmark datasets and develop an interactive and user-friendly web server application based on MSF-HierGNN. Experiments based on benchmark datasets demonstrate that our model not only can effectively extract molecular substructure features and capture correlations among atomic features, but also can more accurately predict molecular properties, thereby offering a novel AI method and application to support drug discovery and design.