TL;DR
This paper introduces HTINet2, a novel deep learning framework for predicting herb-target interactions (HTIs). By integrating Traditional Chinese Medicine (TCM) knowledge with modern biological data through a knowledge graph and employing a residual-like graph neural network, HTINet2 significantly improves the accuracy of identifying potential therapeutic targets for herbs.
What Problem Does This Paper Address?
Identifying the specific biological targets of herbs is a critical but challenging step in drug discovery and development. Traditional methods rely on costly and time-consuming wet experiments. While computational approaches have advanced, existing methods often fail to fully leverage the rich, multi-faceted knowledge embedded in Traditional Chinese Medicine (TCM), such as herb properties, efficacy, and clinical applications. Furthermore, many current models use unsupervised learning, limiting their ability to fully utilize known HTI data for improved prediction accuracy. This paper aims to bridge this gap by developing a supervised deep learning model that integrates diverse TCM knowledge with biological networks for more accurate herb-target prediction.
Methods
HTINet2 is built upon three core modules:
- TCM and Clinical Knowledge Graph Embedding (KGE): The authors construct a comprehensive knowledge graph (TMKG) that fuses TCM properties (like flavor, meridian, efficacy) with clinical treatment information (diseases, symptoms) from various databases and classical texts. They then use knowledge embedding techniques (like DeepWalk, GraRep, node2vec, etc.) to learn low-dimensional vector representations of entities within this graph, capturing their latent relationships.
- Residual-like Graph Representation Learning: A novel residual-like Graph Convolutional Network (GCN) is employed. This network takes the learned entity embeddings as initial features and propagates information through multiple layers to capture complex interactions between herbs and targets within a heterogeneous herb-target network. The "residual-like" aspect, inspired by ResNet, helps mitigate issues like gradient vanishing in deep networks and improves feature aggregation.
- Supervised Target Prediction: A Bayesian Personalized Ranking (BPR) loss function is used for supervised training. BPR is a ranking-based loss that aims to maximize the score of known herb-target pairs (positive samples) while minimizing the score of unknown pairs (negative samples), leading to more accurate predictions.
Datasets
- Herb-Target Interactions (HTI): The primary dataset for HTI was sourced from the SymMap [15] database. After filtering for reliability (inferred evidence score > 0.4616 and p-value < 0.05), a benchmark dataset of 38,002 HTI relationships involving 563 herbs and 2106 targets was established.
- Knowledge Graph Construction: The TMKG was built by integrating data from multiple sources, including:
- SymMap [15]
- soFDA [16] (syndrome ontology)
- STRING [17] (protein-protein interactions)
- KEGG [18] (genomic encyclopedia)
- GO [19] (gene ontology)
- Classical TCM books (e.g., "Pharmacopoeia of the People's Republic of China 2020").
Key Insights and Ideas
- The Power of Integrated Knowledge: The study demonstrates that combining TCM-specific knowledge (properties, clinical uses) with biological network data in a knowledge graph significantly enhances herb-target prediction accuracy. Simply using biological networks is insufficient.
- Residual GCN for Deeper Interactions: The residual-like GCN architecture is crucial for effectively capturing complex, multi-hop interactions within the herb-target network, leading to more robust representations.
- Supervised Learning for Accuracy Boost: Employing a supervised approach like BPR loss, which leverages known HTIs to guide learning, provides a substantial performance improvement over unsupervised methods.
- Interpretability through Visualization: Visualizing the learned embeddings using t-SNE shows that different KE methods cluster entities meaningfully, and HTINet2's learned representations also exhibit good clustering, suggesting a correlation between embedding quality and prediction performance.
Key Results
- State-of-the-Art Performance: HTINet2 achieved significantly better performance than various baseline methods. Specifically, HR@10 increased by 122.7% and NDCG@10 by 35.7% compared to the best baseline (CN).
- Component Contributions: Ablation studies showed that both the knowledge graph embedding and the residual GCN components contribute positively to HTINet2's performance, with the residual GCN having a slightly larger impact.
- Case Study Validation: For Artemisia annua and Coptis chinensis, HTINet2 successfully predicted known targets and identified novel potential targets. Validation through databases (ETCM, HERB), literature, and molecular docking supported the reliability of these predictions. For example, 13 out of 15 predicted targets for Artemisia annua were supported by evidence.
Limitations and Future Work
- Data Noise and Quality: The authors acknowledge that current herb-target relationship data can be noisy. Future work will focus on creating larger, higher-quality datasets by integrating multiple sources and employing advanced techniques for data alignment and semantic disambiguation.
- Two-Stage vs. End-to-End: HTINet2 is a two-stage framework where KE models are pre-trained separately. The authors aim to develop a fully end-to-end framework where KE parameters are optimized jointly with the GCN.
- Interpretability: While visualizations are provided, deep neural networks are often considered "black boxes." Future research will incorporate explainable AI techniques to provide deeper insights into the model's decision-making process.
- Clinical Translation: The authors are committed to accelerating the clinical translation of predicted targets, potentially through wet experiments to validate new discoveries and further refine computational models.
Summary updated