You are here:
Using Chemical Structure Information to Predict In Vitro Pharmacokinetic Parameters (SOT)
Citation:
Pradeep, P., R. Judson, AND G. Patlewicz. Using Chemical Structure Information to Predict In Vitro Pharmacokinetic Parameters (SOT). Presented at SOT 2017 annual meeting, Baltimore, MD, March 12 - 16, 2017. https://doi.org/10.23645/epacomptox.5189371
Impact/Purpose:
Poster presentation at the SOT 2016 annual meeting.Toxicokinetic data are key for relating exposure and internal dose when building in vitro-based risk assessment models. However, conducting in vivo toxicokinetic studies has time and cost limitations, and in vitro toxicokinetic data is available only for a limited set of chemicals. Data gap filling techniques are commonly used to predict hazard in the absence of empirical data. The most established techniques are read-across and quantitative structure-activity relationships (QSARs). This study aims at utilizing both of these techniques to develop predictive models for two in vitro toxicokinetic parameters: fraction unbound (fub) in plasma and intrinsic clearance.
Description:
Toxicokinetic data are key for relating exposure and internal dose when building in vitro-based risk assessment models. However, conducting in vivo toxicokinetic studies has time and cost limitations, and in vitro toxicokinetic data is available only for a limited set of chemicals. Data gap filling techniques are commonly used to predict hazard in the absence of empirical data. The most established techniques are read-across and quantitative structure-activity relationships (QSARs). This study aims at utilizing both of these techniques to develop predictive models for two in vitro toxicokinetic parameters: fraction unbound (fub) in plasma and intrinsic clearance. The analysis relied on a dataset of ~7k chemicals with predicted exposure data, of which 974 chemicals had human in vitro fub in plasma and 540 chemicals had human in vitro intrinsic clearance data. Chemotyper and PubChem fingerprints, and PaDEL descriptors were used as chemical structural features. Unsupervised feature selection was performed to remove the features with less than 80% variance. Read-across and QSAR models were then developed as follows: (1) Read-across: unsupervised KMeans clustering was performed on training chemicals, and was used for predicting clusters for the test chemicals. Next, similarity-weighted read-across predictions were made for each of the test chemicals using analogs from the cluster within a threshold of similarity range. (2) QSAR: The continuous descriptors were normalized to have mean=0 and standard deviation=1. QSAR models were then developed using random forest and support vector machines. The models were evaluated using a 5-fold cross-validation scheme. The best read-across and QSAR model for fub (range: 0 - 1) in plasma had a cross-validated root-mean-squared-error of 0.24 and 0.25, respectively. The difference in parameter mean values between clusters, as measured with a T-test, suggest that fub in plasma is more structurally predictable than intrinsic clearance. Future work includes building predictive models for intrinsic clearance and incorporation of uncertainty in experimental data to quantify uncertainty in model predictions.
URLs/Downloads:
DOI: Using Chemical Structure Information to Predict In Vitro Pharmacokinetic Parameters (SOT)