HTINet2: herb-target prediction via knowledge graph embedding and residual-like graph neural network

Area of Science:

• Pharmacology
• Computational Biology
• Drug Discovery

Background:

• Accurate target identification is vital in drug development for understanding drug mechanisms and discovering new therapeutic targets.
• Existing herb target prediction methods face challenges due to incomplete clinical knowledge and limitations of unsupervised models.

Purpose of the Study:

• To develop an advanced deep learning framework, HTINet2, for accurate herb target prediction.
• To address data and model limitations in current herb target identification approaches.

Main Methods:

• Constructed a large-scale knowledge graph integrating Traditional Chinese Medicine (TCM) properties and clinical treatment data.
• Employed deep knowledge embedding to learn representations of herbs and targets.
• Utilized a residual graph convolution network for interaction learning and Bayesian personalized ranking loss for supervised prediction.

Main Results:

• HTINet2 demonstrated superior performance compared to baseline methods, with a 122.7% increase in HR@10 and a 35.7% increase in NDCG@10.
• Ablation studies confirmed the positive contribution of individual modules within HTINet2.
• Case studies on Artemisia annua and Coptis chinensis validated the reliability of predicted targets using literature and molecular docking.

Conclusions:

• HTINet2 offers a robust deep learning framework for enhancing herb target prediction accuracy.
• The framework effectively integrates diverse knowledge sources and advanced network architectures.
• HTINet2 shows promise for accelerating drug discovery and elucidating herb action mechanisms.