ChronicDPIpredictor: An Interpretable Deep Learning Framework for Chronic and Subchronic Toxicity Assessment
Abstract
Assessing chronic and subchronic toxicity is essential for chemical safety evaluation and risk assessment. ChronicDPIpredictor is a web-based platform that applies fingerprint-based machine learning models to predict compound toxicity from SMILES input.
Compounds are categorized into three toxicity levels based on LOAEL thresholds:
The models demonstrated strong predictive performance. In external validation, three-class classification achieved accuracies of 0.83 for chronic toxicity and 0.81 for subchronic toxicity. When reduced to binary classification (toxic vs. non-toxic), the models reached accuracies of 0.93 and 0.83, with robust sensitivity, specificity, and MCC values.
