Interspecies correlation estimation (ICE) models predict supplemental toxicity data for SSDs
Environmental Protection Agency, USA
Species sensitivity distributions require a large number of toxicity values for a diversity of taxa to define a hazard level protective of multiple species. For most chemicals, measured toxicity data are limited to a few standard test species that are unlikely to adequately represent ecological communities. Interspecies correlation estimation (ICE) models are log-linear least squares regressions that predict the acute toxicity to untested taxa from known toxicity of a single surrogate species. A suite of ICE models is developed from a comprehensive, standardized dataset of acute toxicity with the goal of maximizing the number of potential species for which toxicity can be predicted while minimizing extraneous sources of variation in the models. The United States Environmental Protection Agency houses 3 ICE databases: aquatic animals (vertebrates and invertebrates; 5501 records, 180 species, 1266 chemicals), algae (1647 records, 69 species, 457 chemicals), and wildlife (birds and mammals; 4329 records, 156 species, 951 chemicals). Approximately 2400 models have been developed from these databases and made available through the Web-based Interspecies Correlation Estimation internet application (Web-ICE; http://epa.gov/ceampubl/fchain/webice/).
ICE models were validated using leave-one-out cross validation and sources of model uncertainty were evaluated. Toxicity predictions are most accurate for models with closely related taxa pairs, with over 90% of cross-validated values predicted within 5-fold of the measured value when the surrogate and predicted taxa are in the same family. Model mean square error and prediction confidence intervals should be considered when evaluating an ICE predicted value. Models built with a single mode of action (MOA) were often more robust than models built using toxicity values with multiple MOAs, and improve predictions among species pairs with large taxonomic distance (e.g. within phylum). SSDs developed solely from ICE-predicted toxicity values produce hazard levels with an average factor of 3.0 and 5.0 of those developed with all measured data for aquatic species and wildlife, respectively. For chemicals in which more measured data are available, ICE models may be used to augment datasets to increase species diversity in SSDs. Compared to SSDs developed from only measured data, the uncertainty of ICE model predictions contributes less variability to hazard levels than variance due to species composition. Through extensive study of ICE model evaluation and uncertainty and their application in developing SSDs, ICE generated toxicity values have been demonstrated to provide a statistically sound approach to supplementing datasets to generate SSD-based hazard levels applicable to ecological risk assessments.