Deep learning (DL) offers an opportunity to change the landscapes in toxicity prediction based on structure-activity relationship study in the big data era. The performance of different structural description scenarios in DL models is still controversial. Two structural characterization methods were adopted here to develop respective convolutional neural network models. One used SMILES string for the only input of the DL model, while the other introduced unordered point cloud for chemical structure description. Different DL architectures were exploited to tailor the model structure to the data type. Two DL models on chemical estrogenicity classification were finally obtained for method application. Each model was trained by the chemicals tested in a battery of 18 estrogen receptor (ER)-related assays of the ToxCast program. Both models recognized the active and inactive chemicals at balanced accuracies, which outperformed other machine learning models. Besides, there seems no obvious progress in classification accuracy can be observed for the model using complex input matrix, and the unprecedented learning power of the DL method made the structural characterization much easier. However, more insight into toxicity mechanism can be gained by visualizing critical points and extracting data-driven features of active chemicals. The DL model with point cloud as input even succeeded in recognizing the activity difference of stereoisomers.. Our results demonstrated the necessity of continuous method innovation in SAR.

Deeping learning for enhanced toxicity classification of environmental chemicals: Applications in both SMILES and image-based structure input

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.