RECENT ADVANCES IN AI BASED TOXICITY PREDICTION FOR DRUG DISCOVERY

Toxicity evaluation is a crucial component of drug discovery, as chemical compounds that enhance human health may also pose serious risks. Conventional in vitro and in vivo approaches, while informative, are time-consuming, costly, and dependent on extensive experimentation. The rapid growth of computational approaches and the availability of large experimental datasets have enabled the development of artificial intelligence (AI)–based
toxicity prediction models that provide a more efficient strategy for early-stage screening. By integrating public databases such as ChEMBL, DrugBank, and BindingDB with proprietary in vitro, in vivo, and clinical data, AI models create a feedback loop that improves predictive performance and supports regulatory decision-making. Recent progress in machine learning and deep learning, including Random Forest, Support Vector Machines, Graph Neural Networks, and transformer-based architecture, has demonstrated strong performance in predicting diverse endpoints such as hepatotoxicity, cardiotoxicity, neurotoxicity, and genotoxicity. This article reviews these advances and discusses strategies such as multi-task
learning, multimodal integration, and scaffold- based evaluation to address challenges of data heterogeneity, class imbalance, and protocol variability. Despite these developments, issues
such as data quality, interpretability, and standardization remain barriers to widespread implementation. Overall, this review emphasizes that AI-based toxicity prediction holds great
promise to accelerate drug discovery, reduce attrition, and guide safer therapeutic development.

Keywords: Artificial Intelligence, Machine Learning, Toxicity Prediction, Drug Discovery, Computational Toxicology, Deep Learning, Graph Neural Networks, Regulatory Toxicology.