High Throughput Fate Predictions

High-throughput Simulation of Environmental Chemical Fate for Exposure Prioritization J.F. Wambaugh, D. Reif, S. Gangwal, J. Mitchell-Blackwood, J. Arnot, O. Jolliet, R. Judson, P.P. Egeghy, J. Rabinowitz, D. A. Vallero, W. Setzer, and Elaine Cohen Hubal

International Society for Exposure Science Advancing Exposure Science to Assure Chemical Safety for Sustainability November 1, 2012 Seattle Washington Office of Research and Development

The views expressed in this presentation are those of the author and do not necessarily reflect the views or policies of the U.S. EPA

Introduction There are thousands of chemicals, many without enough data for evaluation Risk is the product of hazard and exposure High throughput in vitro methods beginning to bear fruit on potential hazard for many of these chemicals Methods exist for approximately converting these in vitro results to daily doses needed to produce similar levels in a human (IVIVE)

Judson et al., (2011) “Estimating Toxicity-Related Biological Pathway Altering Doses for Highthroughput Chemical Risk Assessment” Chemical Research in Toxicology 24 451-462

Is exposure science amenable to high throughput methods? How will it be able to keep pace with high throughput biology methods? 1

Office of Research and Development

Oral Equivalent Doses and Estimated Exposures (mg/kg/day)

Oral Doses Equivalent to ToxCast Concentrations

2

Green squares indicate estimated exposures from EPA REDs or CDC NHANES: ~71% of Phase I ~7% of Phase II Office of Research and Development

Wetmore et al. Tox. Sci (2011)

High Throughput Exposure Predictions Goal: A high-throughput exposure approach to use with the ToxCast chemical hazard identification. Proof of Concept: Using off-the-shelf models capable of quantitatively predicting exposure determinants in a high throughput (1000s of chemicals) manner

Environmental Fate and Transport

To date have found only fate and transport models to have sufficient throughput (MitchellBlackwood et al., submitted) These models predict the contribution from manufacture and industrial use to overall exposure We are developing new, high throughput models of consumer use and indoor exposure 3


Consumer Use and Indoor Exposure

Framework for High Throughput Exposure Screening

(Bio) Monitoring

Exposure Inference

Dataset 1

Inferred Exposure

Space of Chemicals (e.g. ToxCast, EDSP21)

Apply calibration and uncertainty to other chemicals

Estimate Uncertainty

Calibrate models

Dataset 2

…

Model 1

Joint Regression on Models

Model 2 4


…

Evaluate Model Performance

USEtox and RAIDAR Treat different models like related high-throughput assays USEtox

United Nations Environment Program and Society for Environmental Toxicology and Chemistry toxicity model Version 1.01 Rosenbaum et al. 2008 5


RAIDAR

Risk Assessment IDentification And Ranking model Version 2.0 Arnot et al. 2006

Parameterizing the Models Cl/C(Cl)=C/C3C(C(=O)OCc2cccc(O c1ccccc1)c2)C3(C)C

Model parameters from EPI Suite, many predicted from structure (SMILES)

Principal component one: half-life in soil, water, and sediment Principal component two: Bioconcentration Factor Principal component three: half-life in air Larger spheres are those for which NHANES biomonitoring data was available 6


Partitioning Release into the Environment Models predictions depend upon release profiles Release profile can be chemical-specific, class-specific, or default depending on data If we have the data then we would use it, but we don’t

Food-use Pesticide Urban Air

USEtox

Rural Air Freshwater

Agricultural soil Natural soil

Sea water

TSCA / Industrial Urban Air

RAIDAR 7

Soil


Freshwater

Agricultural soil Natural soil

Air

Rural Air

Sea water

Air

Water

Soil

Water

…this is much more complicated for indoor/consumer use!

High Throughput Fate Predictions Clustering 1763 chemicals by the media into which they partition most Both models assume exposure scenarios that relate environmental media to food and inhalation exposure Can calculate intake fraction (population exposure in kg per kg emitted – Bennett et al. (2002))

8


Data Availability for Model Predictions and Ground-truthing Ground—truth with CDC NHANES urine data Focusing on U.S. median initially Capable of adding population variability, but will need consumer use models

9


Linking NHANES Urine Data and Exposure Steady-state assumption

(mg/kg/day)i =

parent

metabolite

mg i g creatine 1 * 70 kg g creatine day

Mage et al. (2004), Lakind and Naiman (2008)

Unknowns: 2 (mass, bioavailability) Knowns: 1 (mass)

Unknowns: 2 Unknowns: 3 (fractions), Knowns: 1 (mass balance) Knowns: 3 (mass)

Unknowns: 6 Unknowns: 3 (fractions), Knowns: 3 (mass balance) 10


Knowns: 1 (mass)

Stoichiometry of NHANES Parents and Metabolites

11


Parent chemical Chemical measured In urine

Bayesian Model for fij Further complicated by limit of detection of NHANES chemicals – many chemicals that are checked for are below the LoD However, a finite number of parent exposures are related to a finite number of urine products, and most of relationships are zero, Use Markov Chain Monte Carlo to explore range of possible parent predictions Also incorporate uncertainty in production volume and use all quantiles of NHANES data

Unknown fraction fij for each urine product j due to parent i:

(mg/kg/day) j = MW j ∑ f ij i

12


(mg/kg/day)i MWi

Framework for High Throughput Exposure Screening

(Bio) Monitoring

Exposure Inference

Dataset 1

Inferred Exposure

Space of Chemicals (e.g. ToxCast, EDSP21)

Apply calibration and uncertainty to other chemicals

Estimate Uncertainty

Calibrate models

Dataset 2

…

Model 1

Joint Regression on Models

Model 2 13


…

Evaluate Model Performance

Calibrate ExpoCast Predictions to CDC NHANES Data

Y ~ b f + m f ,1 log(vu ) + m f , 2 log(vr ) + N (bn * + m1,n log(vu ) + m2.n log(vr ) ) Comparison between model predictions and biomonitoring data indicates correlation Indoor/consumer use is critical: Compounds with near-field applications much higher

14


Exposure Prioritization from ExpoCast Uncertainty of prediction indicated by the horizontal confidence interval from the empirical calibration to the NHANES data Five putative use categories – personal care products, consumer use, fragrance, pharmaceuticals, and food additives – were aggregated into a single “near field” indicator variable. 15


Highest Priority

ExpoCast + ToxCast

~93% Phase I and ~89% Phase II coverage by ExpoCast Chemicals with indoor/consumer use in red 16


Chemicals with far field (industrial/agricultural) release in blue

High Throughput Near Field Models • Far field models + near field indicator variable model explains ~10% of NHANES variance • What about purely statistical models with more factors (e.g. fragrance, food additive, octanol:water partition coefficient, vapor pressure)? • We can use eight parameters to explain 60% of variance for NHANES chemicals • Consumer Use • Fragrance • Pharmaceutical • Pesticide • Colorant • Herbicide • Octanol:water partition coefficient • National production volume • For high throughput exposure prediction the models need not be purely mechanistic 17


Exposure Priorities Is exposure science amenable to high throughput methods? How will it be able to keep pace with high throughput biology methods? •

•

•

18

Obtaining new chemical data – Measuring model parameters – Characterizing model appropriateness – Expanding domain of applicability Using existing data – Formatting for computation – Obtaining access Monitoring – Validation of predictions – Characterization of chemical exposure • Specific demographics • Pooled (average) samples Office of Research and Development

•

New indoor/consumer use models

Image from Little et al. (2012), see also Nazaroff et al. (2012), Bennett et al. (2012), Wenger and Jolliet (2012)

Conclusions • High throughput computational model predictions of exposure is possible • These prioritizations have been compared with CDC NHANES data, yielding empirical calibration and estimate of uncertainty • Indoor/consumer use is a primary determinant of NHANES exposure • Developing HT models for exposure from consumer use and indoor environment (post-doc position available) • Use and evaluate these models as additional HT exposure assays • Critical data needs • Measure phys-chem properties outside current domain of applicability • Measure formulation-specific properties • Expand coverage of biomonitoring data

19


U.S. E.P.A. Office of Research and Development ExpoCast Team NCCT John Wambaugh David Dix Alicia Frame Sumit Gangwal Richard Judson Robert Kavlock Thomas Knudsen Stephen Little Shad Mosher James Rabinowitz David Reif Woody Setzer Amber Wang

NERL Peter Eghehy Kathie Dionisio Dan Vallero ARC Arnot Research & Consulting Jon Arnot University of Michigan Olivier Jolliet Michigan State University Jade Mitchell-Blackwood

The views expressed in this presentation are those of the author and do not necessarily reflect the views or policies of the U.S. EPA