Acute toxicity is an important endpoint for chemical risk assessment. As animal testing is still the major avenue for assessing chemical’s adverse effects, in silico screening represents perhaps the most promising alternative approach. Quantitative structure-activity relationships (QSAR) models can be used to estimating chemical induced toxicity with continuous (regression) or discrete (classification) predictions. It is still quite unclear how those different types of models can complement and help each other to afford the best prediction accuracy for a given chemical. Herein, we will discuss a novel, dual-layer hierarchical modeling method to build QSAR models for predicting categorical (binary toxic/nontoxic and four EPA-defined categories) and continuous (LD50) endpoints for assessing rat acute oral toxicity. The training and external test sets were compiled by NICEATM/NCCT and contained 8,213 and 2,843 compounds, respectively. Each molecule has at least one label of ‘toxic/nontoxic’, EPA category (class I, II, III, and IV) and LD50 values. The first layer of independent models (base models) was solely built with computed chemical descriptors. Then, a second layer of models (hierarchical models) were built by stacking all the cross-validated predictions from the base models. In addition, the applicability domain of these models was defined according to the ‘model prediction zone’, which is based on the prediction probabilities of all binary and multiclass models, resulting in four prediction zones: ‘in-zone’ (low confidence), ‘half-zone-binary’, ‘half-zone-multiclass’, and ‘out-zone’ (high confidence). All models were then evaluated using the external test set. The results showed that (i) hierarchical models take advantage of the knowledge learned by the base models, resulting in improved prediction performances (R2 = 0.47, RMSE = 0.62 for base model; R2 = 0.54, RMSE = 0.58 for hierarchical regression model on external test set); (ii) for each individual model, ‘out-zone’ predictions afforded the best prediction performances (up to R2 = 0.65, RMSE = 0.51 for hierarchical regression model with a coverage of 64% on the external test set); and (iii) hierarchical models have larger coverage of ‘out-zone’ predictions than those of the corresponding base models. Overall, this hierarchical H-QSAR modeling method relying on the full stacking of binary, multiclass, and regression base models represents a promising approach for chemical risk assessment.

Hierarchical H-QSAR modeling method that integrates binary/multi classification and regression models for predicting acute oral systemic toxicity

2019 Spring National Meeting

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