ChronicDPipredictor: an interpretable deep learning framework for chemical chronic and subchronic toxicity assessment.

Chronic and subchronic toxicity are very important endpoints for evaluating the long-term and medium-term toxicity of chemical substances. However, due to the complex mechanism and diverse chemical structures, developing effective computer models remains a significant challenge. In this study, we developed ChronicDPipredictor, an interpretable machine learning framework for chemical chronic and subchronic toxicity assessment. Among three fingerprint representations (MACCS, PubChem, and KRFP), models based on MACCS fingerprints achieved the best performance, with accuracies up to 0.82 for chronic and 0.80 for subchronic toxicity in three-class classification. When using these three-class models for the binary classification of toxic and non-toxic substances, the overall prediction accuracy for the chronic toxicity and subchronic toxicity of the compounds reached 0.93 and 0.83 respectively, also showing excellent predictive ability. The SHAP analysis was applied to enhance interpretability. We implemented the ChronicDPipredictor framework as a web-server for predicting the chronic and subchronic toxicity of compounds, which can be freely accessed and used via http://ChronicDPi.sapredictor.cn/ . Furthermore, we systematically extracted structural alerts (18 for chronic toxicity and 7 for subchronic toxicity) from KRFP fingerprints. Several representative alerts, such as nitrobenzene, phenylhydrazine, and triazole derivatives, were supported by established toxicological mechanisms, including oxidative stress, DNA damage, enzyme inhibition, and metabolic disruption. This study provides an interpretable and practical method for assessing chronic and subchronic toxicity, which is quite helpful for the risk assessment of compound repeated-dose toxicity.