Developing an artificial intelligence method for screening hepatotoxic compounds in traditional Chinese medicine and Western medicine combination

Chen Zhao, Mengzhu Zhao, Liangzhen You, Rui Zheng, Yin Jiang, Xiaoyu Zhang, Ruijin Qiu, Yang Sun, Haie Pan, Tianmai He, Xuxu Wei, Zhineng Chen, Hongcai ShangChinese 2022 · DOI

TL;DR

This paper introduces an AI-driven approach to identify compounds that can cause liver damage (hepatotoxicity), particularly when traditional Chinese medicine (TCM) and Western medicine (WM) are combined. By training machine learning and deep learning models on a large dataset of known hepatotoxic compounds, they developed a method that can effectively screen new compounds, potentially improving drug safety and clinical decision-making.

What Problem Does This Paper Address?

Drug-induced liver injury (DILI) is a significant and common adverse drug reaction, often leading to the termination of clinical trials or the withdrawal of drugs from the market. While DILI is a concern for single drugs, the increasing use of combined therapies, especially the integration of Traditional Chinese Medicine (TCM) with Western Medicine (WM), introduces a layer of complexity. This combination, known as TCM-WMC, involves a vast array of compounds, making it challenging to predict and assess the risk of hepatotoxicity. Current methods for identifying hepatotoxic compounds, such as in vitro and in vivo experiments, are often inefficient, costly, and may not perfectly translate to human responses. Therefore, there's a critical need for more effective, scalable, and predictive methods to screen for hepatotoxic compounds within these complex therapeutic regimens.

Methods

The researchers employed a suite of Artificial Intelligence (AI) techniques, specifically focusing on machine learning (ML) and deep learning (DL) models, to develop their hepatotoxicity screening method.

  • Data Preparation: Chemical structures of compounds were represented using SMILES strings. These were then processed using the PaDEL-Descriptor software to calculate a comprehensive set of molecular descriptors and fingerprints, capturing various chemical properties.
  • Machine Learning Models: Nine established ML algorithms were utilized:
  • Stochastic Gradient Descent (SGD)
    • k-Nearest Neighbor (kNN)
    • Support Vector Machine (SVM)
    • Naive Bayes (NB)
    • Decision Tree (DT)
    • Random Forest (RF)
    • Artificial Neural Network (ANN)
    • AdaBoost
    • Logistic Regression (LR)
  • Deep Learning Model: A Deep Belief Network (DBN) was also constructed. DBNs are known for their ability to learn hierarchical representations of data, making them suitable for complex pattern recognition.
  • Training and Testing: The collected dataset was randomly split into training (75%) and testing (25%) sets to evaluate model performance on unseen data.
  • Evaluation Metrics: Model performance was assessed using standard metrics: Classification Accuracy (CA), F1-score, Precision, Recall, and Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve.

Datasets

The study utilized a comprehensive dataset of hepatotoxic compounds derived from multiple sources:

  • Public Databases:
    • DILIrank
    • LiverTox
    • LTKB (Liver Toxicity Knowledge Base)
    • Hepatox
  • Published Literature: Data was also extracted from 11 published DILI-related literature sources.

After data collection and cleaning (removing duplicates, compounds with ambiguous DILI concerns, and those lacking structural information), a final dataset of 2035 hepatotoxic compounds was established. This dataset was then split into a training set of 1505 compounds and a test set of 530 compounds.

Key Insights and Ideas

  • The Power of Ensemble and Deep Learning: While several machine learning models performed reasonably well, the study highlights that deep learning models, specifically the DBN, demonstrated superior performance on the test set. This suggests that the hierarchical feature learning capabilities of deep networks are particularly beneficial for complex tasks like predicting hepatotoxicity from chemical structures.
  • Random Forest as a Strong Baseline: Among the traditional machine learning models, Random Forest (RF) consistently showed strong performance on both training and test sets, often outperforming other ML algorithms. This reinforces RF's robustness and effectiveness in classification tasks involving complex feature spaces.
  • Importance of Chemical Structure Representation: The study underscores the critical role of molecular descriptors and fingerprints derived from chemical structures (SMILES) as input features for AI models. This approach allows for a quantitative and predictive assessment of hepatotoxicity without relying solely on experimental data.
  • TCM-WMC Complexity Demands Advanced Tools: The inherent complexity of combined TCM-WM therapies necessitates sophisticated computational tools like AI to navigate the vast chemical space and identify potential safety risks.

Key Results

  • Overall Dataset Size: A total of 2035 hepatotoxic compounds were collected and used for model development.
  • Random Forest Performance: On the training set, RF achieved an accuracy of 0.838, F1-score of 0.827, Precision of 0.832, and Recall of 0.838. On the test set, RF achieved an accuracy of 0.767 and an AUC of 0.739.
  • Deep Belief Network (DBN) Superiority: The DBN model achieved the highest accuracy on the test set, reaching 82.2%. This performance was noted to be better than all nine other machine learning models tested.
  • Comparative Performance: The paper presents detailed performance metrics for all tested models in Table 1, illustrating the varying degrees of success across different algorithms.

Limitations and Future Work

  • Focus on Single Compounds: The current study primarily focuses on the risk of liver injury caused by single compounds within TCM-WMC. The authors acknowledge that the interaction between multiple compounds in TCM-WMC could lead to synergistic or antagonistic effects on hepatotoxicity, which was not directly modeled.
  • Metabolite and Interaction Effects: Future research should expand to investigate the potential liver injury risks posed by metabolites of TCM-WMC compounds and the complex interactions between these compounds.
  • Mechanism Elucidation: While the AI models can predict hepatotoxicity, further research is needed to explain the underlying mechanisms of liver injury in TCM-WMC.
  • Data Expansion: The authors suggest calculating and incorporating data for more compounds to further refine the predictive models.

Overall, this work provides a valuable foundation for developing AI-driven tools to enhance the safety assessment of TCM-WMC, guiding clinical practice and improving patient safety.

Summary updated

Abstract

The DILI AI models were expected to effectively screen hepatotoxic compounds in TCM-WMC.
BibTeX
@inproceedings{zhao2022developing,
  author = {Chen Zhao and Mengzhu Zhao and Liangzhen You and Rui Zheng and Yin Jiang and Xiaoyu Zhang and Ruijin Qiu and Yang Sun and Haie Pan and Tianmai He and Xuxu Wei and Zhineng Chen and Chen Zhao and Hongcai Shang},
  title = {{Developing an artificial intelligence method for screening hepatotoxic compounds in traditional Chinese medicine and Western medicine combination}},
  booktitle = {Chinese},
  year = {2022},
  doi = {10.1186/s13020-022-00617-4},
}
Figure