Clinical Trial Toxicity Prediction by Using Artificial Neural Network Optimized by Grey Wolf Optimization

Drugs, composed of chemical and biological substances, aim to improve health but often pose toxic risks, including carcinogenic and mutagenic effects. With only 10% of Phase 1 drug candidates gaining FDA approval—mostly due to efficacy and safety issues—traditional drug discovery remains slow, risky, and inefficient, with a 96% failure rate. In silico methods, leveraging machine learning, offer a more efficient approach to predicting toxicity by automating analysis and reducing reliance on experimental trials. However, the effectiveness of these methods is often hindered by the labor-intensive nature of manual parameter tuning, which requires significant expertise and computational resources. This research specifically contributes by integrating Grey Wolf Optimization (GWO) for optimizing Artificial Neural Networks (ANN) in clinical toxicity prediction, offering a novel metaheuristic approach to address the manual parameter tuning problem, which is rarely explored in existing studies. The best-performing ANN configuration, featuring five hidden layers, a tanh activation function, and Adam optimizer utilizing the ClinTox dataset with 1,491 chemical compounds, the optimized ANN achieved 0.966 accuracy and 0.708 F1-Score, indicating high capability to detect toxic compounds in imbalanced datasets.