Using Chemical Structure Information to Develop Predictive Models for In Vitro Toxicokinetic Parameters to Inform High-throughput Risk-assessment

Abstract

The toxicokinetic (TK) parameters fraction of the chemical unbound to plasma proteins and metabolic clearance are critical for relating exposure and internal dose when building in vitro-based risk assessment models. However, experimental toxicokinetic studies have only been carried out on limited chemicals of environmental interest (~1000 chemicals with TK data relative to tens of thousands of chemicals of interest). This work evaluated the utility of chemical structure information to predict TK parameters in silico; development of cluster-based read-across and quantitative structure-activity relationship models of fraction unbound or fub (regression) and intrinsic clearance or Cl_int (classification and regression) using a dataset of 1487 chemicals; utilization of predicted TK parameters to estimate uncertainty in steady-state plasma concentration (C_ss); and subsequent in vitro–in vivo extrapolation analyses to derive bioactivity-exposure ratio (BER) plot to compare human oral equivalent doses and exposure predictions using androgen and estrogen receptor activity data for 233 chemicals as an example dataset. The results demonstrate that fub is structurally more predictable than Cl_int. The model with the highest observed performance for fub had an external test set RMSE/σ=0.62 and R²=0.61, for Cl_int classification had an external test set accuracy = 65.9%, and for intrinsic clearance regression had an external test set RMSE/σ=0.90 and R²=0.20. This relatively low performance is in part due to the large uncertainty in the underlying Cl_int data. We show that C_ss is relatively insensitive to uncertainty in Cl_int. The models were benchmarked against the ADMET Predictor software. Finally, the BER analysis allowed identification of 14 out of 136 chemicals for further risk assessment demonstrating the utility of these models in aiding risk-based chemical prioritization.

1. Introduction

Human health risk assessment associated with environmental chemical exposure is limited by the tens of thousands of chemicals with little or no experimental in vivo toxicity data ¹. The wealth of in vitro toxicity data generated over the last decade has emerged as a promising alternative to animal testing and has enabled better insight into potential mechanism(s) of toxicity ^1-5. However, in vitro toxicity data suffers from a drawback in that it cannot account for the toxicokinetic (TK) factors such as bioavailability, plasma protein binding and intrinsic clearance which are required for the transformation of an in vitro active concentration to a relevant in vivo oral equivalent dose (OED) below which significant in vitro bioactivity is not expected to occur. However, these parameters can be measured, and TK models can be built using them, yielding estimates of steady-state plasma concentration (C_ss). The OED can then be calculated as the ratio of an in vitro potency value (e.g. an AC50) to the C_ss value ^6-10.

Incorporation of toxicokinetic and exposure information can be used in chemical prioritization and can facilitate the addition of a risk context to high-throughput in vitro screening results ^{6-8, 11-13}. Two key experimental TK parameters that are required for relating oral dose to an internal steady state plasma concentration are fraction unbound in plasma (fub) and intrinsic clearance (Cl_int). Although these parameters can be measured experimentally in vitro ^{7, 8, 14}, the protocols are not high-throughput, primarily due to the need to develop chemical-specific analytical methods. As a result, in vitro TK data are available only for fraction of environmental chemicals of interest (~1000 to date), which in turn limit the ability to provide bioactivity exposure ratio (BER) estimates for most environmental chemicals.

In the absence of experimental data, in silico approaches such as read-across ^15-20 and quantitative structure-activity relationship (QSAR) models ^{21, 22} can potentially be used to predict fub and Cl_int. Several in silico models that have been derived for predicting fub^23-29 as well as Cl_int^30-32. Some of these models have been published in the peer reviewed literature, whilst others have been implemented into commercial software tools, such as ADMET Predictor (Simulations Plus Inc., Lancaster, CA). Most of these models were derived using data generated for pharmaceutical chemicals and their relevance for environmental chemicals is unclear.

Here, we derive new in silico models for fub and Cl_int using data extracted from published literature collected for 1486 environmental chemicals ^{7-9, 33}. This study aimed at (1) evaluating the suitability of chemical structure information for predicting these parameters in silico, (2) exploring the utility of read-across and QSAR modeling techniques for developing predictive models for the two in vitro TK parameters, (3) evaluating the implications of variability in experimental and predicted TK parameters, and physicochemical properties on the uncertainty in resultant OED estimates, and (4) integration of IVIVE methods along with high-throughput exposure predictions using the EPAs ExpoCast tool ^{34, 35} to facilitate rapid risk-assessment and chemical prioritization.

2. Workflow

The overall workflow in this study comprised three main steps (Supplemental Figure S1). First, experimental data along with fingerprints and molecular descriptors were used to develop QSAR models. Second, the predictions from the models developed in this work were compared with the predictions from the commercially available ADMET Predictor package. Last, the predictions from this work were used to calculate OEDs using IVIVE methods implemented in the HTTK package and compared with the human exposure predictions from ExpoCast ^{34, 35} to facilitate high-throughput risk-assessment. The methods are described in detail in the following section. Additional analysis was performed where unsupervised clustering of chemicals with human in vitro TK parameters data was used to determine whether structurally related chemicals had similar TK parameters. The clusters derived using unsupervised clustering were used along with experimental data to derive cluster-based read-across predictions. These analyses do not directly impact the main findings of this study and are discussed in supplemental information (1-Supplemental-Methods&Results.docx).

3. Methods

3.1. Dataset

The data used in this analysis was obtained from published literature and available through the high-throughput toxicokinetic (HTTK) R package ^{7, 33, 36-47}. The dataset consists of 1486 chemicals that span a variety of use classes including pharmaceuticals, food-use chemicals, pesticides and industrial chemicals ⁴⁸ of which 1139 chemicals had experimental human in vitro fub data and 642 chemicals that had experimental human hepatic in vitro Cl_int data. An external dataset of 1,814 chemicals tested in a battery of 18 ER and 11 AR related assays ^{49, 50} was also utilized in this study for model validation. Before developing any models, the chemicals from this external ER-AR dataset were removed to ensure there is no training bias in the predictions for external validation on this set. All the structures were curated and the sdf file format was obtained from the DSSTox database ^{51, 52}. The distribution of experimental values for fub and Cl_int are shown in Supplemental Figure S2. Since the data were non-normally distributed, they were appropriately transformed before any analysis was conducted. The details of the transformation and the transformed data distribution are presented in the results section and Supplemental Figures S2 and S3. A complete list of chemicals with CAS registry numbers (CASRN) and experimental data for both parameters and the chemical structure sdf file are included as supplemental information (2-fub_data.csv, 3-clint_data.csv, and 5-QSARreadyStructures.sdf).

3.2. Molecular descriptors

The chemicals used in this study were characterized using two structure-based fingerprints PubChem ⁵³ and ToxPrint chemotypes ⁵⁴; two physicochemical descriptors (acid dissociation constant, acidic and basic pKa and logarithm of water-octanol partition coefficient, logP) computed using the OPERA software ⁵⁵; 12 molecular descriptors calculated using the Chemistry Development Kit (CDK) ^{56, 57} implemented in KNIME the KNIME analytics platform ⁵⁸ (version 2.11.3); and 1875 descriptors (1444 1D, 2D and 431 3D descriptors) calculated using PaDEL software⁵⁹. PubChem fingerprints were generated in the KNIME analytics platform ⁵⁸ (version 2.11.3). ToxPrints were generated within the publicly available Chemotyper application (version1.0.r12976, https://chemotyper.org). The fingerprint and descriptor generation require an sdf file format which was obtained from the DSSTox database ⁶⁰. The final descriptor selection and preparation for both fub and Cl_int datasets was done as follows:

1. PubChem fingerprints and ToxPrints were combined to generate one combined fingerprint. Feature selection was performed to remove features with less than 80% variation across the chemical set, and only one feature from a pair of features was retained if the Pearson correlation coefficient between them was more than 80%,

2. The OPERA pKa acidic (pKa_a) and pKa basic (pKa_b) predictions were used to infer a pKa value as the lower of the 2 values. In case, only one them exists then that was used as the pKa,

3. All continuous descriptors were normalized to have mean = 0 and standard deviation = 1,

4. A supervised recursive feature elimination algorithm was used to select 10 descriptors from PaDEL and CDK descriptors combined, and

5. All chemicals for which the fingerprints/descriptors could not be calculated were dropped from the analysis.

A complete list of all the fingerprints and descriptors used in the final models is provided in table S6 of supplementary information (1-Supplemental-Methods&Results.docx).

3.3. QSAR modeling

3.3.1. Data Preparation

Fraction unbound in plasma

Chemicals with a fub value equal to 0 (below limit of detection) were set at a default of 0.005, based on the assumptions in the HTTK package ⁴⁷, and those with a value of 1 were set to 0.99 (upper limit of detection). The final dataset after removing the chemicals from the external ER-AR dataset comprised 1003 chemicals that had defined Pubchem fingerprints and Toxprints. Before modeling, fub was transformed using the log-odds ratio form ²⁵:

$$fu{b}_{transformed}=lo{g}_{10}\frac{(1-fub)}{fub}$$ (2)

Supplemental Figure S3 shows the distribution of the transformed fub values. This transformed dataset was divided into an 80% training dataset (802 chemicals) and a 20% external test dataset (201 chemicals). Regression QSAR models were developed for fub, details of which are provided below.

Intrinsic Clearance

The final dataset after removing the chemicals from the external ER-AR dataset comprised 524 chemicals that had defined Pubchem fingerprints and Toxprints. Each chemical was assigned a clearance value of low, medium or high based on the following scheme: Cl_int values less than or equal to 0.9 were set as low, values between 0.9 and 50 were set as medium and values greater than 50 were set as high. These thresholds were defined by Ekins et al. ³⁰. All Cl_int values are in units of uL/min/million cells. Two types of QSAR models were developed for Cl_int: (1) classification models to predict low, medium and high clearance, and (2) regression models using data for chemicals in the medium clearance group to avoid the effect of outliers (low or high clearance values) on the models. For regression models, the clearance data from the medium bins was log transformed before modeling. Supplemental Figure S4 shows the distribution of clearance data for classification and regression models. For both classification and regression models, the transformed dataset was then divided into two parts where 80% of the data (419 chemicals for classification and 269 chemicals for regression) was used as the training set and 20% of the data (105 chemicals for classification and 68 chemicals for regression) was used as an external test set.

Once the classification (low, medium, high) and regression (medium bin) QSAR models for intrinsic clearance were run, the results were combined as follows. Chemicals predicted to have low clearance were assigned a default clearance value equal to the median clearance value for training set chemicals in the low clearance bin, and the chemicals predicted to have high clearance were assigned a default value equal to the median clearance value for training set chemicals in the high clearance bin. Thus, the models can predict if a chemical exhibits low, medium or high clearance, and provides a quantitative value in each case.

3.3.2. Algorithm

Regression models for fub and Cl_int were developed using the lasso regression ⁶¹, support vector machine (SVM) ^{62, 63}, random forest (RF) ^{64, 65} and neural network multiple layer perceptron ^{65, 66} algorithms. Classification models for Cl_int were developed using the logistic regression ^{65, 67}, support vector machine (SVM) ^{62, 63}, random forest (RF) ^{64, 65} and neural network multiple layer perceptron ^{65, 66} algorithms. The datasets were randomly split into a training set (80% chemicals) and an external test set (20% chemicals). The training set was used to build the models using 5-fold cross-validation with hyper-parameter tuning where the model is developed over a grid of parameter values and the values with the best model performance are selected as the final algorithm parameters. The final models were then evaluated on the external test set. Detailed discussion of the machine learning algorithms and the hyperparameters tuned for each model are available in supplemental information (1-Supplemental-Methods&Results.docx).

For both fub and Cl_int, several models were developed in an additive fashion where a new set of descriptors were added incrementally to observe any improvements in model predictivity. The first set of models, referred to as the baseline models, were developed using structural information encoded by PubChem fingerprints and ToxPrints. The subsequent models expanded on the baseline models in terms of the descriptor space. In the second set of models, two physicochemical descriptors (LogP and pKa) were added to the fingerprints. In the third set of models, 10 additional physicochemical descriptors from PaDEL and CDK were added. Finally, the two best performing models for each endpoint were combined to develop a consensus model ⁶⁸, where the final prediction was the average prediction from the two best performing models. The performance of each regression model was evaluated in terms of MAE (mean absolute error), RMSE (root mean square error), RMSE/σ (RMSE / standard deviation of the endpoint distribution) and the variance explained (R²). The performance of each classification model was evaluated in terms of accuracy, F1-score (harmonic mean of positive predictive value and sensitivity) for each class and R². Different additive models were compared based on the improvement in performance relative to their coverage. Note that the number of chemicals used in developing each additive model was dependent on generation of valid set of descriptors for each model.

3.4. Prediction of TK parameters and comparison with ADMET predictor

Model predictions from this work were compared with those of the ADMET Predictor package on an external dataset of 1,814 chemicals tested in a battery of 18 ER and 11 AR related assays ^{49, 50}. ADMET predictions were obtained for these chemicals using data previously generate by Sipes et al ¹². Overall, 472 chemicals had predicted fraction unbound and 410 chemicals had predicted Cl_int from the models developed in this work and the ADMET predictor. The residuals (difference between the observed experimental value and the predicted value) from both models (this work and ADMET predictor) were compared across both TK parameters to identify any chemicals or regions where both the models performed poorly. The experimental data for selected chemicals was then re-evaluated to account for potential experimental errors and data anomalies. Note that the chemicals in this set were not used in training the models in this work but it is unknown if those chemicals were present in the training dataset for the ADMET predictor.

3.5. Prediction and validation of C_ss: Calculation of in silico C_ss and comparison with in vitro C_ss

The HTTK R package ⁴⁷ was used to predict steady-state concentration in plasma (C_ss) using the R software environment ⁶⁹. C_ss is interpreted as the steady-state concentration of a chemical in the plasma given a constant 1 mg/kg/day oral dose rate and has units of μM/mg/kg/day. The default parameters of the population simulator in HTTK, httkpop (using calc_mc_Css) were used ^{47, 70} to calculate C_ss. The httkpop function returns the upper 95^th percentile of C_ss in the population – corresponding to individuals for whom the same 1 mg/kg/day exposure produces plasma concentrations higher than 95% of the population. This is intended to be a conservative estimate, calculated with a simple steady-state model that estimates clearance from passive renal filtration and well-stirred hepatic metabolism. The omission of other routes of clearance acts to make C_ss higher (more conservative when comparing to estimated exposure). The calculation of C_ss for each chemical requires chemical-specific physicochemical properties (LogP and pKa), in addition to fub and Cl_int. Experimental fub and Cl_int data was available for 709 chemicals from the HTTK package. So, in vitro C_ss (using experimental fub, Cl_int and the default physicochemical properties from the HTTK package) and in silico C_ss (using predicted fub, Cl_int and the physicochemical properties calculated from the OPERA tool) could be calculated for 709 chemicals. The HTTK package LogP was obtained from the DSSTox database ⁵¹ or predicted using the EPA’s estimation program interface (EPI) suite (http://www.epa.gov/tsca-screening-tools/epi-suitetmestimation-program-interface) or using OPERA⁷¹. pKa was taken from the literature when available or taken from predictions produced by Strope et al ⁷². The in vitro and in silico C_ss calculations were compared to get an estimate of variance in C_ss values based on the models developed in this work.

3.6. Comparison of human oral equivalent doses (OEDs) and exposure predictions: Bioactivity-exposure ratio plot

A set of OEDs for each chemical in the dataset described in Section 3.4 was calculated by dividing in vitro potency values (ACC: activity concentration at cut-off) for a series of in vitro assays by the corresponding C_ss values. The assays that were used measured activities in the estrogen and androgen receptors (ER and AR), and have been combined to provide definitive estimates of agonist and antagonist activity for these receptors ^{50, 73}. The OEDs for each chemical were then compared with the median estimated daily exposure taken from the EPAs ExpoCast estimates ^{34, 35}. For each chemical, the following analysis was done:

1. The response (ACC value) in any ER or AR related assay was retained if the agonist or antagonist model AUC value ^{50, 73} was greater than 0.1, indicating activity in the specific target and mode (ER, AR, agonist, antagonist). The lowest ACC value across all assays (the most potent assay) was then used to calculate a conservative estimate of OED and is referred to as ACC hereafter,

2. The ACC was divided by the chemical’s in silico C_ss value to obtain an OED,

3. If the chemical had an in vitro C_ss value, ACC was divided by the in vitro C_ss value to obtain an overall conservative estimate of OED for the chemical based on in vitro C_ss,

4. An estimate of variance in the in silico values based on the C_ss prediction analysis from the analysis in Section 3.5 was incorporated to obtain an overall conservative estimate of OED for the chemical based on in silico C_ss. The ACC was divided by the in silico C_ss plus twice the standard deviation of the residuals between in silico and in vitro C_ss values,

5. Finally, all estimates of OEDs were compared with the exposure estimates.

The software for data analysis and model development was developed using the functions implemented in the scikit-learn module ⁷⁴ of Python 2.7 ⁷⁵ and is available as supplemental information (code.zip).

4. Results and Discussion

4.1. QSAR modeling

Fraction Unbound in Plasma

Feature selection on combined PubChem fingerprints and Toxprints resulted in 80 substructural features that were used for baseline model development. Subsequent models expanded the baseline feature set (80 features) with additional physicochemical descriptors. Across all sets of models, the best predictive performance was achieved when using RF, SVM or Lasso algorithms. Consequently, the consensus models were developed by averaging the predictions of best two models. Supplemental Table S2 summarizes the model details and the performance metrics for all the models developed. The coverage for all the baseline models was 1003 chemicals (the entire dataset). Adding two physicochemical descriptors LogP and pKa(from OPERA) to the baseline models resulted in the best consensus model performance metrics with MAE = 0.61, RMSE = 0.82, RMSE/σ = 0.66 and R² = 0.56 for 5-fold internal cross-validation, and MAE = 0.61, RMSE = 0.80, RMSE/σ = 0.65 and R² = 0.57 for the external test set validation with over 10% loss in coverage at 886 chemicals. Figure 1 shows the observed versus predicted transformed Fub values as evaluated in 5-fold internal cross-validation (red dots) and external validation (blue squares). Note that the RMSEs for all the models are well within one standard deviation (=1.24) of the endpoint distribution (Figure S3), which provides a context to the error rates.

Figure 1.

Observed versus predicted fraction unbound (transformed scale) for the random forest and support vector machine consensus model (highlighted in red in Supplemental Table S2) for 5-fold internal cross-validation (red dots) and external test set validation (blue squares). The root-mean-squared-error and R-squared for the 5-fold internal cross-validation are 0.82 and 0.56, respectively, and for external test set validation are 0.80 and 0.57, respectively. The black solid line indicates the line of perfect fit, where the predicted values would equal the experimental values. The red dashed lines indicate an error margin of ±1 standard deviation of the training dataset and the blue dotted lines indicate an error margin of ±1 standard deviation of the test dataset. The uncertainty in the observed data is indicated as a red error bar (coefficient of variation for uncertainty = 0.4) on an example chemical.

Intrinsic Clearance

Feature selection on PubChem fingerprints and Toxprints resulted in 79 substructural features that were used for baseline model development. Subsequent models expanded the baseline feature set (of 79 features) with additional physicochemical descriptors. As expected from the unsupervised clustering analysis results (see supplemental file Supplemental-Methods&Results.docx for the methods used), use of structural descriptors did not yield highly predictive models. Supplemental Tables S4 and S5 summarize the model details and the performance metrics for all the classification and regression models, respectively. Adding LogP and pKa as descriptors to the classification baseline models using the support vector algorithm resulted in the best consensus model performance metrics with accuracy = 66.47% and F1-score = [0.43, 0.77, 0.18] for 5-fold internal cross-validation, and accuracy = 73.26% and F1-score = [0.41, 0.83, 0.25] for the external test set validation. The regression models were all poorly performing with the random forest baseline model being the best one with MAE = 0.39, RMSE = 0.46, RMSE/σ = 1.01 and R2 = −0.02 for 5-fold internal cross-validation, and MAE = 0.33, RMSE = 0.41, RMSE/σ = 0.92 and R2 = 0.14 for the external test set validation. Figure 2 shows the observed versus predicted transformed clearance values as evaluated in 5-fold internal cross-validation (red dots) and external test set validation (blue squares).

Figure 2.

Observed versus predicted medium intrinsic clearance (transformed scale) for the random forest model (highlighted in red in Supplemental Table S5) for (a) 5-fold internal cross-validation (red dots), and (b) external test set validation (blue squares). The root-mean-squared-error and R-squared for the 5-fold internal cross-validation are 0.46 and −0.02, respectively, and for external test set validation are 0.41 and 0.14, respectively. The black solid line indicates the line of perfect fit, where the predicted values would equal the experimental values. The red dashed lines indicate an error margin of ±1 standard deviation of the training dataset and the blue dotted lines indicate an error margin of ±1 standard deviation of the test dataset. The uncertainty in the observed data is indicated as a red error bar (coefficient of variation for uncertainty = 0.3) on an example chemical.

Overall, fub models performed better than Cl_int models. Unsupervised clustering analysis (described in supplemental file 1-Supplemental-Methods&Results.docx) show that: (i) fub values are more tightly bounded across different clusters as compared to Cl_int, and (ii) the mean value of fub are more distinct across clusters as compared to Cl_int. A new evaluation of uncertainty in experimental Cl_int values identified a median coefficient of variation of 0.31 [Wambaugh 2019, submitted]. Therefore, there is significant uncertainty in the experimental values on which the QSAR modeling was performed.

4.2. Prediction of TK parameters and comparison with ADMET predictor

Figure 3(a) shows the plot of the residuals between the current model and those of ADMET Predictor for fraction unbound for 472 chemicals. The RMSE and R² for ADMET Predictor were 0.19 and 0.55, respectively, and the RMSE and R² for this work were comparatively better for this dataset at 0.16 and 0.67, respectively. The residuals from the two models had low correlation (R² = 0.40). In general, the ADMET predictor over-predicts fub while the consensus model from this work under-predicts fub as compared to the experimental data. The dotted lines and the dashed lines highlight the chemicals for which the absolute residuals were greater than 0.25 and 0.50 for both predictors (outliers), respectively. The outliers with experimental data from Wetmore et al. 2012 and 2015 were cross-checked for any potential measurement errors^{7, 46}. In general, the experimental data appeared to be reliable with the predicted fraction unbound values lower than the measured data. It is speculated that the ADMET predictor looks specifically at predicted binding to AAG and albumin, but does not consider additional binding to lipoprotein complexes, whereas the experimental measures capture these interactions.

Figure 3.

Residual comparison plots between the models developed in this work with those of the ADMET predictor. The black line on the plots is the line of perfect fit, where the residuals (observed minus predicted) from both the predictors are the same. (a). Plot of fraction unbound residuals. The dotted lines highlight the regions where the absolute residuals were greater than 0.25 for both the predictors. The dashed lines highlight the regions where the absolute residuals were greater than 0.50 for both the predictors. (b) Plot of intrinsic clearance residuals. The dashed lines highlight the region where the residuals were less than −2 log-units for both the predictors. The dotted line highlights the region where the residuals from the ADMET predictor were greater than 3 log-units.

Figure 3(b) shows the plot of the residuals for Cl_int for 410 chemicals on the log₁₀-transformed dataset. For the purposes of plotting on a logarithmic-scale, the chemicals with an observed clearance value of zero were defaulted to 10⁻³ (i.e. log-transformed clearance value = −3). The models developed in this work worked better on this dataset with explained variance of 0.22 as compared to the negative R² values from the ADMET predictor. However, the RMSEs were 1.89 and 1.55 for ADMET and this work, respectively. The residuals from the two models were quite uncorrelated (R² = 0.17). The dashed lines highlight the region where the residuals for both predictors were less than −2 log-units and dotted line highlights the region where the residuals from the ADMET predictor were greater than 3 log-units (outliers). In general, the residuals from this work are more centered around zero, indicating lesser bias, as compared to the ADMET predictor which has a wider range of residuals. The outliers with experimental data from Wetmore et al. 2012 and 2015 were cross-checked for any potential measurement errors (details described in the supplementary file 1-Supplemental-Methods&Results.docx)^{7, 46}. In general, the discrepancies in primary hepatocyte data can be accounted due to non-P450 related clearance (ADMET predictor uses only P450 metabolism in their calculations), non-metabolic degradation and sensitivity/detection issues at 1μM measurements. The poor predictions from the models developed in this work could be due to the general inability of chemical structural descriptors to adequately model Cl_int. A complete list of all the chemicals highlighted in the residual plots, experimental and predicted data, and data evaluation results are provided in supplemental information (4-ERAR-FubClintData.xlsx).

4.3. Prediction and validation of Plasma C_ss: Calculation of in silico C_ss and comparison with in vitro C_ss

Figure 4 shows a plot of C_ss values calculated using the models in this work (in silico C_ss) against C_ss calculated using the data in HTTK package (in vitro C_ss) for the ER-AR dataset and the entire set of chemicals with in vitro fub and clint data from the HTTK package. The C_ss units are log10 mg/kg. The observed RMSE and the R² values for the ER-AR dataset were 0.82 and 0.47, and for the big dataset were 0.83 and 0.40, respectively. These plots incorporate the variability in C_ss calculations owing to the underlying experimental variability in the fub and Cl_int data as calculated using the HTTK package and are shown as error bars in the plot. The results of this analysis demonstrate that C_ss calculations tend to be relatively stable given the uncertainty in (predicted) clearance values. The standard deviation of the residuals from this analysis were further used in the BER analysis to arrive at a conservative estimate of dose for hazard assessment. Note that the analysis excludes all chemicals that were in the training dataset for development of the fub and Cl_int models.

Figure 4.

Comparison of C_ss calculated using the models in this work (in silico C_ss) calculations with C_ss calculated using the data in HTTK package (in vitro C_ss) for (a) the ER-AR dataset and (b) the entire set of chemicals with in vitro fub and clint data from the HTTK package. The plots incorporate the variability in C_ss calculations owing to the underlying experimental variability in the fub and Clint data. The standard deviation of residuals from this analysis were used in the BER analysis (for the ER-AR dataset) to derive a conservative estimate of dose for hazard assessment. The C_ss units are log₁₀ mg/kg. Note that the analysis excludes all chemicals that were in the training dataset for development of the fub and clint models.

4.4. Comparison of human oral equivalent doses (OEDs) and exposure predictions: Bioactivity-exposure ratio plot

Figure 5 shows the range of OEDs derived for each chemical, based on ER-AR in vitro assay data, using C_ss values calculated based on the fraction unbound and clearance predictions from this work, along with exposure predictions. A total of 136 chemicals were active for ER or AR and had C_ss predictions. The OED value from the lowest assay potency (ACC) and using the in silico C_ss estimate is represented as a blue square dot. The exposure predictions are represented as green squares with error bars indicating the 95% upper confidence interval of the median exposure estimate. The red solid circles indicate OEDs derived using in vitro C_ss values but are only shown for chemicals that had in vitro fraction unbound and clearance data. The black squares indicate a conservative estimate of OEDs derived using in silico C_ss predictions and incorporating the standard error (=2σ of the residuals) due to uncertainty in C_ss predictions for the ER-AR dataset (Figure 4(a)).

Figure 5.

Bioactivity-exposure ratio (BER) plot to compare human oral equivalent doses (OEDs) and exposure predictions. Lowest and a conservative estimate of ToxCast derived OEDs (visualized as blue and black squares, respectively) were compared with the exposure estimates (visualized as median exposure value along with a 95% upper confidence interval in green). The plot is ordered by lowest ToxCast OED estimate. This analysis allows for the generation of a BER plot that compares hazard to exposure estimates within a high-throughput risk assessment framework to aid chemical screening and risk-prioritization.

About 92% (121/136) of the chemicals have a hazard point estimate (lowest OED value from in vitro assays, blue squares) higher than the upper 95% confidence interval (CI) of the exposure estimate. After adjusting for uncertainty, (black squares), about 91% (120/136) of chemicals have an OED higher than the upper 95% CI exposure estimate. For chemicals with in vitro OED estimates, 85% (120/136) have an OED higher than the upper 95% CI exposure estimate, leaving 16 that would be prioritized for follow-up based on an overlap between the OED and the exposure estimate. Of these, the in silico estimates identified 14 of the 16. In general, the chemicals on the left side of the plot with overlapping exposure and OED estimates are naturally occurring hormones or pharmaceuticals that are intended to occur at levels that are bioactive. A complete list of chemicals and the corresponding data are available in supplemental information (3.ERAR_BERPlot-Data.csv).

Overall the novelty of this work lies in (1) the development of open-source QSAR models for fub and Cl_int using a simple descriptor space and a rich chemical dataset, allowing prediction of pharmacokinetics for thousands of chemicals for which experimental data is not available, (2) incorporation of uncertainty due to the source of physicochemical properties in C_ss calculations in the estimation of OEDs to allow for a conservative comparison with exposure predictions resulting in higher confidence, and (3) extending the ability to prioritize large numbers of data-poor chemicals using in silico predictions in an effort to ease the transition from hazard-based prioritization to exposure related risk assessment. The model of fub shows higher predictivity than does that for Cl_int, but the Cl_int model developed here has similar predictivity to the commercial model it was compared to. The model for C_ss combines fub, Cl_int and physicochemical parameters, which are also typically the result of QSAR models. A recent analysis using new experimental fub and Cl_int measurements characterizes the uncertainty in the observed data and estimates the coefficient of variation for uncertainty as 0.4 for fub and 0.3 for Cl_int data [Wambaugh submitted 2019]. Both this new analysis and that shown here indicates that C_ss is less sensitive to uncertainty in Cl_int than to the other inputs (fub, pKa, logP). A final important point is illustrated by Figure 5 and the associated discussion. The uncertainties associated with using the in silico C_ss values are on the same order of magnitude as the uncertainties in the exposure predictions, but the BER (minimum OED, uncertainty included/maximum exposure) are often large compared with the individual component uncertainties. This means that in general, classifications of chemicals into those with BER > or < 1 are relatively accurate. 16 of 18 chemicals with exposure ratio<1 using in vitro TK parameter values were identified with in silico TK parameter values. This indicates that the models developed here for fub and Cl_int can provide useful input for efforts to prioritize thousands of chemicals lacking the appropriate experimental data.