Herb Target Prediction Using Machine Learning Methods in a Heterogeneous Network

Predicting herb-target interactions is crucial for advancing traditional Chinese medicine (TCM), but the existing methods often struggle with incomplete datasets and fail to fully leverage the network structure.   The objective of this study is to develop a novel network-based approach that integrates both network topology as well as molecular data to further improve the accuracy of herb-target interaction predictions. We have constructed a heterogeneous network encompassing herbs, targets and symptoms, incorporating various network measures to assess edge weights. Molecular data for herbs has also been integrated into the model. Using six ensemble supervised learning models (GBM, XGBoost, LightGBM, CatBoost, etc.),  the model has been trained to predict herb-target interactions. The proposed  model achieved an AUROC of 88% and an AUPR of 90% on the HIT2 dataset, significantly outperforming the existing approaches. This research highlights the potential of integrating network structure and molecular data for accurate herb-target prediction, opening new avenues for drug discovery and personalized medicine in TCM.