You are here:
Use of Supervised Learning Approaches with Hierarchical Fragment-based Representations (ToxPrints) to Characterize Mode-of-Action in Aquatic Fish Toxicity
Citation:
Chang, D., K. Fay, K. Mansouri, K. Markey, J. Prindiville, G. Patlewicz, M. Lewis, M. Shobair, E. Saluck, AND A. Richard. Use of Supervised Learning Approaches with Hierarchical Fragment-based Representations (ToxPrints) to Characterize Mode-of-Action in Aquatic Fish Toxicity. SERMACS, Durham, NC, October 25 - 28, 2023. https://doi.org/10.23645/epacomptox.24790593
Impact/Purpose:
N/A
Description:
Chemical categorization, or grouping, is routinely employed to capture and report salient chemistry and toxicity correlations as well as to consider analogs for chemicals that have limited empirical information. One prominent application of chemical grouping used by regulatory scientists is the Ecological Structure Activity Relationship (EcoSAR) model, which predicts the toxicity of classes of chemicals to various aquatic species. There is a continual need to update chemical categories and predictive models as new information on chemical hazards becomes available, especially given that a large proportion of industrial substances are not classifiable by EcoSAR and similar tools. This study employed machine learning approaches to evaluate whether potential refinements could be made to the current EcoSAR classes using chemical fingerprint and in vitro biological activity information. Refinements included building sub-categories for broad, existing EcoSAR classes (e.g., neutral organics), as well as identifying new categories to address substances currently unclassifiable by EcoSAR. An ensemble tree-based binary classification model was developed to predict narcotic or specific-acting aquatic toxicity modes of action. The model was trained on chemical fingerprints (i.e., ToxPrints) and/or in vitro biological activity (i.e., EPA’s Toxicity Forecaster, or ToxCast, and the federal research collaboration Tox21, or Toxiciology in the 21st Century, high-throughput screening data). It was then used to predict aquatic toxicity mode of action (narcosis vs specific-acting) for chemicals classified as neutral organics or unclassifiable by the current version of EcoSAR. Chemotype and activity enrichments for those chemicals predicted to be specific-acting identified features useful for refining EcoSAR classes, including several bond chemotypes (e.g., sulfonyl, sulfide, sulfonate, alkyl-tri-halo, and benzopyran) and in vitro assay activity (e.g., Novascreen ENZ assays). This approach identified data gaps in the biological activity inventory, potential analogs for chemicals that may fit the suggested new categories and suggests specific high-throughput assays that may be most useful for informing reductions in animal testing. This abstract does not necessarily represent the views or policies of the U.S. Environmental Protection Agency.
URLs/Downloads:
DOI: Use of Supervised Learning Approaches with Hierarchical Fragment-based Representations (ToxPrints) to Characterize Mode-of-Action in Aquatic Fish Toxicity
PRESENTATION.PDF (PDF, NA pp, 1305.5 KB, about PDF)