Identification of intrinsic hepatotoxic compounds in Polygonum multiflorum Thunb. using machine-learning methods

Area of Science:

• Pharmacology
• Computational Chemistry
• Toxicology

Background:

• Polygonum multiflorum (PM) is a traditional Chinese medicine with known liver benefits but also associated with drug-induced liver injury (DILI).
• The intrinsic hepatotoxicity of natural products within PM (NPPM) is a significant factor in herb-induced liver injury.
• Identifying specific NPPM responsible for DILI is crucial for ensuring the safety of herbal medicines.

Purpose of the Study:

• To identify NPPM with high DILI potential using machine learning (ML) methods.
• To develop and validate a computational approach for predicting hepatotoxicity of natural compounds.
• To investigate the intrinsic DILI mechanism of PM.

Main Methods:

• Collected 197 NPPM and a DILI-labeled dataset of 2384 compounds.
• Employed a diparametric optimization method to tune extended-connectivity fingerprints (ECFPs), Rdkit, and atom-pair fingerprints with machine learning algorithms.
• Validated predictions using K-means clustering and in vitro cell-viability assays with HepaRG cells.

Main Results:

• An optimized Support Vector Machine (SVM) model using ECFPs achieved 0.761 accuracy and 0.834 recall.
• In silico screening identified 47 NPPM with high DILI potential, clustered into six groups.
• Dianthrones and anthraquinones were identified as high-risk compounds, with dianthrones showing the lowest IC50 values (0.7-0.9 μM).

Conclusions:

• Machine learning and in vitro screening successfully identified dianthrones and anthraquinones as major contributors to PM-induced DILI.
• The developed diparametric optimization method is a powerful tool for screening toxicants in large datasets.
• The study provides a validated computational approach for predicting hepatotoxicity of natural products.