You are here:
Informing the Human Plasma Protein Binding of Environmental Chemicals by Machine Learning in the Pharmaceutical Space: Applicability Domain and Limits of Predictability
Ingle, B., R. Tornero-Velez, J. Nichols, AND B. Veber. Informing the Human Plasma Protein Binding of Environmental Chemicals by Machine Learning in the Pharmaceutical Space: Applicability Domain and Limits of Predictability. Journal of Chemical Information and Modeling. American Chemical Society, Washington, DC, 56(11):2243-2252, (2016).
The National Exposure Research Laboratory (NERL) Computational Exposure Division (CED) develops and evaluates data, decision-support tools, and models to be applied to media-specific or receptor-specific problem areas. CED uses modeling-based approaches to characterize exposures, evaluate fate and transport, and support environmental diagnostics/forensics with input from multiple data sources. It also develops media- and receptor-specific models, process models, and decision support tools for use both within and outside of EPA.
The free fraction of a xenobiotic in plasma (Fub) is an important determinant of chemical adsorption, distribution, metabolism, elimination, and toxicity, yet experimental plasma protein binding data is scarce for environmentally relevant chemicals. The presented work explores the merit of utilizing available pharmaceutical data to predict Fub for environmentally relevant chemicals via machine learning techniques. Quantitative structure-activity relationship (QSAR) models were constructed with k nearest neighbors (kNN), support vector machines (SVM), and random forest (RF) machine learning algorithms from a training set of 1045 pharmaceuticals. The models were then evaluated with independent test sets of pharmaceuticals (200 compounds) and environmentally relevant ToxCast chemicals (406 total, in two groups of 238 and 168 compounds). The selection of a minimal feature set of 10-15 2D molecular descriptors allowed for both informative feature interpretation and practical applicability domain assessment via a bounded box of descriptor ranges and principal component analysis. The diverse pharmaceutical and environmental chemical sets exhibit similarities in terms of chemical space (99-82% overlap), as well as comparable bias and variance in constructed learning curves. All the models exhibit significant predictability with mean absolute errors (MAE) in the range of 0.10-0.18 Fub. The models performed best for highly bound chemicals (MAE 0.07-0.12), neutrals (MAE 0.11-0.14), and acids (MAE 0.14-0.17). An RF model constructed with 10 descriptors had the highest accuracy across both pharmaceuticals (MAE 0.146-0.157) and environmentally relevant chemicals (MAE 0.103-0.130). The inclusion of the majority of the ToxCast test sets within the AD of the RF model, coupled with high prediction accuracy for these chemicals, indicates the model provides a QSAR for Fub that is broadly applicable to both pharmaceuticals and environmentally relevant chemicals.