Comparative Evaluation and Data Analysis for Drug Toxicity Prediction

Throughout the entire drug development process, toxicity prediction is a significant process in assessing the possible toxicity of drugs. Recently, there has been active research on estimating drug toxicity using deep learning techniques. However, challenges such as a lack of labeled data and the insufficient reliable benchmark datasets remain major obstacles. To address these challenges, this study analyzes and compares seven toxicity benchmark datasets from the Therapeutics Data Commons and three external literature datasets. Experimental results demonstrated that deep learning methods showed superior performance on the benchmark datasets. In particular, on the ClinTox dataset, deep learning models outperformed conventional machine learning methods by 5.5%. The results confirm that employing self-supervised learning effectively mitigates the data limitation problem. Additionally, it was validated that using graph representation learning and natural language processing effectively handles chemical structures as drug features.