Hepatotoxicity Modeling Using Counter-Propagation Artificial Neural Networks: Handling an Imbalanced Classification...

Area of Science:

• Pharmacology
• Computational Toxicology
• Artificial Intelligence

Background:

• Drug-induced liver injury (DILI) poses significant challenges in drug development.
• Current in vitro and in vivo methods are often complex, costly, and may not fully capture DILI.
• In silico approaches offer a cost-effective alternative for toxicity prediction.

Purpose of the Study:

• To evaluate counter-propagation artificial neural networks (CPANNs) for classifying imbalanced DILI datasets.
• To develop a predictive model for drug hepatotoxicity using CPANNs.
• To optimize CPANN model performance for DILI prediction.

Main Methods:

• Utilized counter-propagation artificial neural networks (CPANNs) optimized with a genetic algorithm.
• Employed molecular descriptors for classifying drugs into hepatotoxic and non-hepatotoxic categories.
• Modified the CPANN training algorithm by incorporating random subsampling to handle imbalanced data.

Main Results:

• The modified CPANN training algorithm with random subsampling demonstrated improved classification ability on imbalanced DILI datasets.
• Models trained using an imbalanced set with balanced subsampling per epoch outperformed those trained on a fixed balanced set.
• Internal validation and external set prediction statistics supported the efficacy of the proposed subsampling approach.

Conclusions:

• Random subsampling integrated into CPANN training is an effective strategy for improving the prediction of drug-induced liver injury from imbalanced datasets.
• The developed CPANN model shows promise for predicting the hepatotoxic potential of drugs.
• This in silico approach offers a valuable complement to traditional experimental methods in drug safety assessment.