In silico prediction of chemical reproductive toxicity using machine learning

• Published in: Journal of applied toxicology 2019-06, Vol.39 (6), p.844-854
• Main Authors: Jiang, Changsheng, Yang, Hongbin, Di, Peiwen, Li, Weihua, Tang, Yun, Liu, Guixia
• Format: Article
• Language: English
• Subjects: Biocompatibility; Biomedical materials; Chemicals; Embryos; Frequency analysis; Gonadal dysgenesis; Hazard assessment; Health hazards; Health risks; In vivo methods and tests; Learning algorithms; Machine learning; molecular fingerprint; Molecular modelling; Organic chemistry; Prediction models; reproductive toxicity; structural alerts; structure‐activity relationship; Substructures; Teratogenicity; Toxicity
• ISSN: 0260-437X; 1099-1263
• DOI: 10.1002/jat.3772

Reproductive toxicity is an important regulatory endpoint in health hazard assessment. Because the in vivo tests are expensive, time consuming and require a large number of animals, which must be killed, in silico approaches as the alternative strategies have been developed to assess the potential reproductive toxicity (reproductive toxicity) of chemicals. Some prediction models for reproductive toxicity have been developed, but most of them were built only based on one single endpoint such as embryo teratogenicity; therefore, these models may not provide reliable predictions for toxic chemicals with other endpoints, such as sperm reduction or gonadal dysgenesis. Here, a total of 1823 chemicals for reproductive toxicity characterized by multiple endpoints were used to develop structure‐activity relationship models by six machine‐learning approaches with nine molecular fingerprints. Among the models, MACCSFP‐SVM model has the best performance for the external validation set (area under the curve = 0.900, classification accuracy = 0.836). The applicability domain was analyzed, and a rational boundary was found to distinguish inaccurate predictions and accurate predictions. Moreover, several structural alerts for characterizing reproductive toxicity were identified using the information gain combining substructure frequency analysis. Our results would be helpful for the prediction of the reproductive toxicity of chemicals. Binary classification models were developed by six machine learning approaches with nine molecular fingerprints aiming at predicting the potential reproductive toxicity (RT) of compounds. Among the models, MACCSFP‐SVM model performed best. The applicability domain of the prediction models was analyzed, and a rational boundary was found to distinguish inaccurate predictions and accurate predictions. Moreover, several structural alerts for characterizing RT were identified using the information gain combining substructure frequency analysis.