Assessment of predictive models for estimating the acute aquatic toxicity of organic chemicals

Abstract

In silico toxicity models are critical in addressing experimental aquatic toxicity data gaps and prioritizing chemicals for further assessment. Currently, a number of predictive in silico models for aquatic toxicity are available, but most models are challenged to produce accurate predictions across a wide variety of functional chemical classes. Appropriate model selection must be informed by the models’ applicability domain and performance within the chemical space of interest. Herein we assess five predictive models for acute aquatic toxicity to fish (ADMET Predictor™, Computer-Aided Discovery and REdesign for Aquatic Toxicity (CADRE-AT), Ecological Structure Activity Relationships (ECOSAR) v1.11, KAshinhou Tool for Ecotoxicity (KATE) on PAS 2011, and Toxicity Estimation Software Tool (TEST) v.4). The test data set was carefully constructed to include 83 structurally diverse chemicals distinct from the training data sets of the assessed models. The acute aquatic toxicity models that rely on properties related to chemicals’ bioavailability or reactivity performed better than purely statistical algorithms trained on large sets of chemical properties and structural descriptors. Most models showed a marked decrease in performance when assessing insoluble and ionized chemicals. In addition to comparing tool accuracy and, this analysis provides insights that can guide selection of modeling tools for specific chemical classes and help inform future model development for improved accuracy.