HightoLow Dose Extrapolation: Issues and Approaches

1
High-to-Low Dose ExtrapolationIssues and
Approaches

• BOSC Workshop on EPA Chemical Risk Assessment
  Principles and Practices
• February 2, 2005
• Weihsueh Chiu, Ph.D.
• National Center for Environmental Assessment,
  U.S. EPA

2
Summary

• EPA uses a variety of approaches for low-dose
  extrapolation.
• A number of important issues need to be
  considered in the choice and implementation of
  these approaches.
• EPA is actively working to
• Advance the science and methods for low-dose
  extrapolation.
• Make use of the best science that is available
  for its current risk assessments.

3
EPA often needs to estimate risks at exposures
below the range of data
Low Dose High Dose
Exposures/Doses of Regulatory Interest
Environmental Epidemiology Exposure Biomarker
Studies
Fewer studies
Occupational Epidemiology Human Pharmacokinetic
Studies
Animal Bioassays Animal Pharmacokinetic Studies
??
In Vitro studies Emerging Data (-omics)
4
Low-dose extrapolation has both biological and
statistical components

• Population dose-response (thick blue lines)
  combines individual/biological dose-response
  (thin lines) and inter-individual variability
  (different thin lines).

5
Approaches to extrapolation below the range of
observation

• Model-independent
• Linear from Point-of-Departure (POD)
• RfD/RfC from POD using uncertainty factors (UFs)
• Model-dependent
• Empirical models (e.g., multistage model)
• Biologically-based models
• Combination of approaches
• Linear from POD for target dose metric,
  pharmacokinetic model for exposure dose metric
• RfD/RfC from POD with data-/ model-based UFs

6
Characteristics of model-independent approaches

• Separation of observed range and extrapolation
  range
• Explicit avoidance of quantifying relationships
  below range of observation
• Choice of linear and non-linear depends on
  knowledge of mode-of-action
• Linear extrapolation from POD generally
  interpreted as a plausible upper bound on
  potency
• RfD/RfC extrapolation from POD defined as likely
  to be without an appreciable risk
• Need for consistency in procedures and results

7
Where observation ends and extrapolation
begins

• Delineation of Point-of-Departure
• Evolution from LOAEL/NOAEL to modeling based on
  observed dose-response (Benchmark Dose)
• Benchmark Dose Software (BMDS) facilitates
  consistency and reproducibility
• Current internal efforts aimed at improving
  consistency in choice of benchmark response
• BMDS currently being expanded to include
  time-to-tumor modeling

8
Model-independent approach depends on knowledge
of mode-of-action

• Linear

• Non-linear

9
Linear extrapolation consistent with previous LMS
method

• Analysis by Subramaniam et al 2005 of LMS and
  linear extrapolation procedures
• Of 102 data sets from IRIS database
• 84 had differences of lt 10 in potency estimates
  from the two approaches
• 95 had lt 2-fold differences.
• Linear extrapolation based on upper bound (LED)
  and MLE (ED) were also compared 75 of datasets
  had lt 2-fold differences.

10
Re-examination of RfD methodology

• Currently completing review of the scientific
  foundations of RfD low-dose extrapolation process
• Data that has been cited as basis for UFs and/or
  developing UF distributions
• Probabilistic methods that have been proposed for
  combining UFs
• See need for a common conceptual framework
• For instance, SAB Panel on TCE (2002) wrote
• Ultimately, the whole system of uncertainty
  factors could be usefully revisited and defined
  in terms of an object of achieving x level of
  risk for the yth percentile of the variable human
  population with z degree of confidence.

11
Characteristics of model-dependent approaches

• Presumption that model (e.g., Mode-of-action,
  structural assumptions) is valid below range of
  observation.
• Often involve unobserved parameters estimated by
  fitting models to dose-response data.
• Can be implemented so as to be interpreted as a
  central estimate (with uncertainty bounds),
  assuming model is true.
• Can be used in combination with model-independent
  approaches

12
Historical Example Empirical cancer
dose-response models

• Well known that different empirical models may
  fit experimental data equally while differing by
  orders of magnitude at low doses
• Food Safety Council (1980)
• Crump and Howe (1985)
• Many others
• Default approach at the time was to use the
  linearized multistage (LMS) model (q1)

13
Example PBPK model uncertainty

• Most PBPK models provide central tendency
  estimates, but little or no characterization of
  uncertainty.
• Tetrachloroethylene (PERC)
• Three human PBPK models with similar model
  structures, but different parameters
• Trichloroethylene (TCE)
• Two different model structures, each with two
  alternative parameterizations

14
PERC Three human PBPK models with different
low-dose predictions

• Human in vivo data weakly constrain metabolism
• Model predictions and observation of parent
  compound blood concentrations within factor of
  2.
• Up to 10-fold range of predicted total metabolism
  in observed range.
• Ten-fold range of extrapolations of metabolism to
  low dose (lt 1 ppm).

15
Example PBPK-based UF for Human Variation

• Ongoing efforts evaluating use of PBPK models to
  help inform choice of toxicokinetic portion of
  UFH
• Chloroform
• MTBE
• TCE
• One implementation issue is accounting for
  residual uncertainty (model or parameter)

16
Chloroform Data on human variability input to
PBPK model

• Impact on level of metabolites of
  inter-individual variability in
• Blood Solubility
• Hepatic Blood Flow
• Enzyme content
• In addition, for children
• Use of age-specific organ sizes and blood flow
• Development of PBPK model for neonates (1 yr) and
  juveniles (9 yr)

17
Example Two-stage (MVK) carcinogenesis model

• Formaldehyde model (Conolly et al 2003, 2004)
• CFD dosimetry model
• Hybrid PBPK-CFD model (and data) for DPX
• Two-stage clonal growth model for nasal cancer
• Parameter sources
• in vitro measurements (e.g., cell labeling)
• Fitting to time-to-tumor data

Normal Cells
n
b
Initiated Cells
a
Death/ Differentiation
m
Malignant Cells
18
Clonal growth models exhibit strong parameter
sensitivity

• Crump (1994) showed examples where low-dose
  extrapolation of MVK model could change by gt105
  with 1 change in initiated cell birth/death
  rates (a or b)
• More generally, low-dose extrapolation thus
  needs
• Reliable information on biological parameters
  and/or their relationships at (low) dose.
• Understanding of the sensitivity of low-dose
  extrapolation to parameter uncertainty and
  variability
• Characterization of the range of risk estimates
  from different plausible model structures.
• In the case of formaldehyde, EPA is working to
  better characterize these and other uncertainties
  and so as to evaluate plausible bounds on risk.

19
Issues for Implementation of Model-Dependent
Approaches

• Characterization of both qualitative and
  quantitative uncertainty / variability.
• Model structure uncertainty, including
  dose-response of model parameters
• Parameter uncertainty, and variability
• Data reliability / relevance
• Ultimate impact on quantitative risk estimate
• Given such a characterization, what level of
  confidence is necessary to replace estimates
  based on model-independent approaches?

20
Summary of EPAs approach to high-to-low dose
extrapolation

• Both model-independent and model-dependent
  approaches used in EPAs current and future risk
  assessments.
• Major issues with choosing and implementing
  different approaches include
• Knowledge of mode-of-action, biological
  relationships at low dose
• Characterization of uncertainty and variability
• Degree of confidence and consistency in results
• EPA is working both to advance the science and
  methods in this area as well as to make use of
  the best science that is available for its
  current risk assessments.

21
Acknowledgements

• NCEA Leadership
• Peter Preuss, Director NCEA
• David Bussard, Director NCEA-W
• Paul White, Chief, QRMG

• NCEA Staff
• Chao Chen
• Karen Hogan
• Jennifer Jinot
• John Lipscomb
• Cheryl Siegel Scott
• Ravi Subramaniam