An Effects-based Expert System to Predict Estrogen Receptor Binding
Affinity for Food Use Inert Ingredients and Antimicrobial Pesticides: 

Application in a Prioritization Scheme for Endocrine Disruptor Screening

NOTICE

THIS DOCUMENT IS A DRAFT

7/27/2009

U.S. Environmental Protection Agency

Office of Pesticide Programs

Washington DC 20460

Mid-Continent Ecology Division

National Health and Environmental Effects Research Laboratory

Office of Research and Development

Duluth MN 55804



PREFACE

The Office of Pesticide Programs has a number of ongoing activities to
advance computational tools to better predict potential hazards and
improve the effectiveness of risk assessment and risk management by
focusing the generation of new data on likely risks of concern.  The use
of quantitative structure-activity relationships (QSARs) in early hazard
identification is one such cost effective prioritization tool that can
guide the systematic collection of key test data.  The Food Quality
Protection Act of 1996 requires EPA to screen all pesticidal chemicals,
which includes active ingredients and other (pesticidally “inert”)
ingredients, for potential endocrine disruption (ED).  The Agency must
prioritize hundreds of chemicals for ED screening and testing in
biologically complex and resource intensive assays. A particularly
challenging issue is the development of prioritization techniques for
chemicals, such as inert ingredients and some antimicrobial active
ingredients that have minimal existing toxicological information.  This
Federal Insecticide Fungicide and Rodenticide Act Scientific Advisory
Panel (FIFRA SAP) review will focus on a (Q)SAR approach that was
developed for pesticide inert ingredients and antimicrobial pesticides
to help prioritize candidate chemicals for Tier 1 screening under the
Endocrine Disruptor Screening Program.  This (Q)SAR rule-based expert
system for predicting estrogen receptor (ER) binding affinity was
developed by the Office of Research and Development in collaboration
with the Office of Pesticide Programs.  

As part of the Organization for Economic Cooperation and Development
(OECD) activities to increase the regulatory acceptance of (Q)SAR
methods, this (Q)SAR model conforms to an “OECD Principles for the
Validation, for Regulatory Purposes, of (Q)SAR Models” (see  HYPERLINK
"http://www.oecd.org/dataoecd/33/37/37849783.pdf"
http://www.oecd.org/dataoecd/33/37/37849783.pdf ) (OECD, 2004a) and was
the subject of the OECD Expert Consultation to Evaluate an Estrogen
Receptor Binding Affinity Model for Hazard Identification (OECD, 2009a).
The Agency has considered and incorporated the OECD comments into the
current white paper for the FIFRA SAP review. The comments and
recommendations agreed to at the OECD Expert Consultation were: 

a. While not all the data [was] reported there is a strong experimental
basis for the system with even more experimental evidence for the
pathway to be presented in the near future.

b. Grouping of chemicals using the proposed expert system looks
promising and provides the most transparent mechanistic explanation
system to date.

c. As an expert system it can be expanded as described in the
consultation document for broader regulatory domains.

d. The tool was developed to be consistent with Level 1 of the OECD
Endocrine Disruption Testing and Assessment Conceptual Framework.

e. The tool is useful to build chemical categories.

f. ER-binding will not automatically translate to reproductive toxicity.
Therefore, there would be greater risk assessment relevance if the
system was assessed for its ability to use ER-binding affinity
predictions to predict effects at higher levels of biological
organization. 

g. The expert system on ER-binding affinity should be automated.

h. It is recommended that the tool should be implemented in the OECD
(Q)SAR Application Toolbox. See
(http://www.oecd.org/document/54/0,3343,en_2649_34379_42923638_1_1_1_1,0
0.html)

i. Further discussions are necessary on how to implement such an expert
system in the Toolbox.

Once the Agency receives and considers the FIFRA SAP comments and
recommendations, it intends to make this model publicly available
through the OECD (Q)SAR Application Toolbox, an international means of
making (Q)SAR technology readily accessible and transparent for use by
risk assessors, as well as through other means.



DISCLAIMER

	None of the discussions surrounding any individual chemical substance
in this paper, EPA presentations at the SAP meeting, or the meeting
summary are intended to support any conclusions other than to aid in the
development of potential priority-setting methods.  None of the
discussions, conclusions or examples presented should be taken as a
conclusion regarding the potential risks of the substance or that the
Agency has or will make a determination that any use of the substance
necessarily will pose a significant risk. Further, the research results
described are not designed to provide a scientific basis, by themselves,
to allow EPA to conclude that these substances are "endocrine
disruptors."  Measured or predicted ER binding affinity data for
substances mentioned in this paper may prove to be useful in developing
future prioritization scheme(s) for EDSP Tier 1 screening.  

	This document has been reviewed in accordance with U.S. Environmental
Protection Agency policy. Mention of trade names or commercial products
do not constitute endorsement or recommendation of use. 

Table of Contents		

							    				 Page

Chapter I.   Prioritizing Chemicals for screening and Assessment: an
Application for the Endocrine Disruptor Screening Program	1	      

Background	1

Prioritizing Testing and Assessment of Pesticide Chemicals	3

Prioritizing Chemicals for Endocrine Disruptor Screening	6

Evolution of the Expert System 	11

Chapter II. Mechanistic Basis Of The Expert System To Predict Relative
Estrogen Receptor (ER) Binding Affinity	12

 

Overview of the Approach to Build an Effects-Based Expert System	12

ER Binding Affinity: An Indicator of Potential Reproductive Effects	14

The ER Binding Domain	16

The Regulatory Chemical Domain	17

Characterization of Inert and Antimicrobial Inventories	18

The Chemical Inventory Domains of Interest and Existing Information	18 

The Receptor Binding Assay Domain	19

Chapter III.   Regulatory Inventories Of Interest And Expert System
Modeling Domain	21

Overlaying Model Domains and Regulatory Inventories	21 

Expanding the ER Binding Affinity Knowledge Base 	23

A Rule-Based Expert System to Predict ER Binding Affinity	30

Chapter IV.  Accessibility of the Expert System	36

Expert Systems for ER Binding Affinity	36

Ongoing Research: Human ER Binding Affinity and Gene Activation	37

Chapter V. Conclusions	39

References	43

Figures & Tables	48

Chapter I.   Prioritizing Chemicals for Screening and Assessment: an
Application for the Endocrine Disruptor Screening Program

	This chapter briefly reviews national and international efforts to
formulate approaches for prioritizing large chemical inventories for
toxicity screening, testing, and risk assessment. The development and
application of structure activity relationships (SARs) and quantitative
structure activity relationships (QSARs) to prioritize inert ingredients
used in pesticide formulations that require a tolerance or tolerance
exemption under the Federal Food, Drug, and Cosmetic Act (FFDCA)
(referred to as “food use inerts” or “FI”) and active
ingredients contained in antimicrobial pesticides (referred to as
“antimicrobial pesticides” or “AM”) for screening in US EPA’s
Endocrine Disruptor Screening Program (EDSP) is then introduced. 

Background

	Hazard assessment of chemicals has historically relied on empirical in
vivo data from specialized test guidelines that are designed to identify
adverse outcomes in different species.  The pace of conducting hazard
assessments is generally limited by the availability of test data for
many environmental chemicals.  There are more new chemicals being
produced each year than can practically be assessed annually for hazards
through the development and review of new in vivo test data which is
both a time and resource intensive process. This trend is likely to
continue unless a more strategic approach is adopted to set priorities
for testing and assessment so the chemicals with greatest potential for
impacts in ecosystems and human populations are evaluated in a timely
manner.  Even for the endpoints covered in the Screening Information
Data Sets (SIDS) established by the Organization of Economic Cooperation
and Development (OECD), which is summary information for commonly used
hazard endpoints (  HYPERLINK "http://www.chem.unep.ch/
irptc/sids/oecdsids/"  www.chem.unep.ch/ irptc/sids/oecdsids/
sidspub.html), there are reasons to lessen the dependence on in vivo
testing all chemicals for all endpoints. These reasons include the
length of time to undertake the studies and generate the data, the large
numbers of animals sacrificed in the testing, as well as the high cost
of toxicity testing and review of the resultant data. These concerns are
“amplified” given that when considering large chemical inventories
and the breadth of SIDS endpoints only a small percentage of compounds
are likely to be identified to pose a significant hazard in any given
test guideline. Consequently, a strategic approach is needed to set
priorities for more focused in vivo testing. A number of recent reports
reflect the growing recognition that more efficient testing strategies
and prioritization approaches are needed in both ecological and human
health risk assessments (Bradbury et al., 2004; NRC, 2007; OECD, 2008). 

	Broadly, there are two primary, and complementary, approaches to
setting priorities for initial hazard identification of untested
chemicals.  One approach is to develop faster high throughput in vitro
screens. Enormous efforts are underway to develop and apply these
methods (e.g., http://www.epa.gov/ncct/toxcast).  The other,
complementary, approach is to develop SARs and QSARs to estimate
relevant endpoint values to form prioritization schemes.  The OECD
member countries have long recognized the potential of (Q)SAR to
establish appropriate priorities for the conduct of initial hazard
assessments for thousands of untested chemicals.  The OECD countries,
with substantial US participation, have developed principles to guide
the validation of (Q)SAR models for regulatory purposes (OECD, 2004a)
and have compiled case studies to illustrate how member countries are
using, or anticipate using, (Q)SAR methods within national regulatory
contexts. 

Prioritizing Testing and Assessment of Pesticide Chemicals

US EPA’s Office of Pesticide Programs (OPP) is committed to protecting
public health and the environment through application of the latest
scientific tools to increase the reliability and effectiveness in
assessing and managing potential pesticide risks (will be available at
http://www.epa.gov/pesticides/science/index.htm). OPP’s critical path
focuses on fully utilizing an integrated approach to testing and
assessment. This vision is consistent with the Toxicology Testing in the
21st Century report published by the National Research Council (NRC,
2007).  The future goal is to move toward a hypothesis-based paradigm
where in vivo testing is targeted to the most likely hazards of concern.
This progressive, tiered-testing approach is envisioned to start with
hazard-based hypotheses about the plausible toxicological potential of a
pesticide or group of pesticide chemicals. Existing exposure and
toxicity information would then be combined with computer modeling and
diagnostic in vitro assays to target toxicity testing to the specific
data needed for human health and ecological risk assessments. These
advances will be incorporated within a risk assessment framework of
problem formulation, hazard, dose response, exposure assessment, and
risk characterization to support pesticide registration decisions. This
understanding will allow the program to advance the risk assessment
approach from one reliant on extensive in vivo animal testing to one
that is based on a chemical’s mode of action (i.e., understanding how
chemicals perturb normal biological function) and a better understanding
of real world exposures to pesticides that is more directly relevant to
humans and wildlife. Achieving these objectives will require close
collaboration with the scientific community, international
organizations, and government partners to build the foundation for
understanding chemically-induced toxicity pathways.  

An initial focus of the Program’s efforts to develop and apply an
integrated approach to testing and assessment is to better determine and
prioritize what toxicity data are needed to inform risk assessments for
chemicals that do not have extensive in vivo toxicity information (e.g.,
inert ingredients and certain antimicrobial pesticides, as well as
metabolites and degradates of pesticide active ingredients).  The
transition to this approach includes maximizing use of existing data
from similar compounds, hazard and exposure modeling, and in vitro data
to prioritize specific animal toxicity testing that is needed to assess
and subsequently manage risks. The Program is including use of
integrated approaches to testing and assessment to prioritize
information needs either across specific chemical inventories, or
endpoints for a given chemical by combining different types of existing
information on a similar chemical or group of structurally similar
compounds, using predictive computer modeling (e.g., SARs/QSARs) and in
vitro and high throughput screening assays.  

The USEPA Office of Pollution Prevention and Toxics (OPPT) has long used
SAR/QSAR to identify and prioritize information needs and estimate
potential hazards of industrial chemicals.  USEPA’s Pesticide Program,
in collaboration with OPPT, has applied these methods in situations
where there is limited data.  In the assessment of inert ingredients,
for example, SARs have been used in identifying structurally-similar
compounds to form categories of chemicals.  This use of SAR allows for
the testing of selected members of a category with the results then
being applied to the other members of the group. Additionally SAR
principles are used to identify chemical “analogs” (chemicals that
are closely related to the chemical of interest) for compounds where
there is limited data on the chemical of interest but for which more
robust data are available on the analog chemical.

The applications of SARs/QSARs are being advanced in both ecological and
human health risk assessments of pesticide inert ingredients.  In the
area of ecological risk assessment, QSAR programs such as ASTER (USEPA,
2009a) and ECOSAR (USEPA, 2009b) are utilized to estimate toxicity to
fish, invertebrates, and algae.  In the area of human health risk
assessment, SAR estimations are used as part of an overall weight of the
evidence approach to better characterize potential risk concerns for
pesticide metabolites and contribute to decisions regarding additional
data requirements.  For example, SAR models for prediction of
carcinogenicity, such as Oncologic (USEPA, 2005a), along with toxicity
data from studies of similar chemicals can help better characterize
uncertainties regarding potential carcinogenicity of a chemical for
which there are no actual cancer studies.  Additionally, analog data can
facilitate the identification of potential toxicological endpoints of
concern and thereby target testing needs or follow-up actions.

 Antimicrobial active pesticide ingredients generally have less data
available for risk assessment than conventional pesticides.  Consistent
with the advancement of SAR/QSAR approaches for industrial chemicals and
pesticide inert ingredients, the Pesticide Program recently released a
White Paper on “Use of Structure-Activity Relationship (SAR)
Information and Quantitative SAR (QSAR) Modeling for Fulfilling Data
Requirements for Antimicrobial Pesticide Chemicals and Informing EPA’s
Risk Management Process” in support of a proposed rule on data
requirements for antimicrobial pesticides. (USEPA, 2008a, Docket:
2008-0110; USEPA, 2008b,  Document 0045;  HYPERLINK
"http://www.regulations.gov/fdmspublic/ContentViewer?objectId=0900006480
665444&disposition=attachment&contentType=pdf"
http://www.regulations.gov/fdmspublic/ContentViewer?objectId=09000064806
65444&disposition=attachment&contentType=pdf )

	The current white paper describes a (Q)SAR-based expert system to
support determinations of the order in which chemicals will be subject
to Endocrine Disruptor Tier 1 screening.  In other words, for
prioritization of food use inert ingredients and antimicrobial
pesticides for screening of endocrine disruptor potential based on their
potential to bind to the estrogen receptor (ER).  This expert system
provides a new tool for the Pesticide Program to advance its integrated
approach to testing and assessment.

Prioritizing Chemicals for Endocrine Disruptor Screening

	The international community, through the OECD Endocrine Disruption
Testing and Assessment Advisory Group (EDTA), identified the great
importance of and urgent need for relatively cheap and quick screens,
and tests not requiring animals, to prioritize chemical evaluation for
potential endocrine disruption. The EDTA subsequently created the
Validation Management Group-Non-Animal (VMG-NA), whose main objective is
to identify or propose validated or promising non-animal assays for
endocrine testing, and in particular to develop and validate tools
necessary for Level 1 (sorting and prioritization with existing data
and/or (Q)SAR systems) and Level 2 (in vitro assays providing
mechanistic data) within the EDTA Conceptual Framework for the Testing
and Assessment of Endocrine Disrupting Chemicals.  Levels 3, 4, and 5 of
the OECD Framework include in vivo testing and assessment for single
endocrine mechanisms and effect, multiple endocrine mechanisms and
effects, and endocrine plus other mechanisms, respectively (OECD,
2004b).  

	In the US, the Food Quality Protection Act (FQPA) mandates the USEPA to
determine whether pesticide chemicals (active and inert ingredients) may
have an effect similar to an effect produced by naturally occurring
estrogen or other endocrine effects. In developing the Endocrine
Disruptor Screening Program (EDSP) the Agency received expert advice
from Federal Advisory Committees, including the Endocrine Disrupter
Screening and Testing Advisory Committee (EDSTAC) and USEPA’s Science
Advisory Board (SAB) and the FIFRA Scientific Advisory Panel (SAP).  The
Agency, consistent with EDSTAC, SAB and SAP input has been developing
the EDSP program that initially includes estrogen, androgen and thyroid
hormone systems and includes an evaluation of potential effects in
humans and wildlife (i.e., fish, birds and amphibians)
(http://www.epa.gov/scipoly/oscpendo/index.htm). Also consistent with
recommendations from the advisory panels, the EDSP encompasses
non-pesticide chemicals, in addition to active and inert ingredients in
pesticide formulations.

	   In its final report, the EDSTAC (  HYPERLINK
"http://www.epa.gov/scipoly/oscpendo/pubs/" 
http://www.epa.gov/scipoly/oscpendo/pubs/  edspoverview/finalrpt.htm)
(USEPA, 1998a) recommended a tiered approach for detecting chemicals
with endocrine disrupting potential using a resource-efficient manner
that is similar to that established by the OECD.  The framework proposed
by the EDSTAC involved: 1) sorting and prioritizing chemicals; 2) Tier 1
screening; and 3) Tier 2 testing.  Tier 1 screening data were envisioned
to provide information to determine if a chemical had the potential to
interact with the estrogen, androgen and/or thyroid systems, while Tier
2 test data would be used to support hazard and risk assessments. This
framework was subsequently adopted in large part by the Agency (USEPA,
1998b) and generally supported by a joint committee of the SAB/SAP
(USEPA, 1999). 

	With regard to sorting and prioritization, the EDSTAC (USEPA, 1998a)
recommended  that chemical inventories be sorted into 4 categories: 1)
chemicals considered unlikely to interact with hormone systems, such as
certain polymers and strong mineral acids and bases; 2) chemicals
without sufficient data to make a determination whether or not to
proceed to Tier 2 testing or hazard/risk assessment; 3) chemicals with
sufficient existing data to make a determination to proceed to Tier 2
testing (i.e., existing data that provided information consistent with
the Tier 1 screening assays); and 4) chemicals with sufficient existing
information to support a hazard assessment.  Given that the vast
majority of chemicals in commerce would fall into category 2, the EDSTAC
provided detailed recommendations for prioritizing chemicals for Tier 1
screening.  The priority-setting framework included the use of existing,
or predicted, exposure and effects information.  To the extent possible,
prioritization was recommended to involve a scheme that combined
exposure and effects information.  With regard to effects information,
the EDSTAC encouraged the Agency to evaluate the use of high-throughput
screening (HTPS) assays for receptor binding and transcriptional
activation and (Q)SARs for these endpoints to obtain empirical or
predicted information on chemicals for which no data were available to
support prioritization.

	The Agency subsequently formulated a framework generally consistent
with the EDSTAC’s recommendations (USEPA, 1998a), which was presented
to the SAB/SAP
(http://yosemite.epa.gov/sab/sabproduct.nsf/C8ABD410E357DBCF85257193004C
42C4/$File/ec13.pdf) (USEPA, 1999).  The SAB/SAP supported sorting
chemicals into the 4 categories and stressed the importance of having
exposure and effects information to prioritize chemicals in category 2.
The SAB/SAP, however, concluded at that point in time HTPS and (Q)SARs
were not sufficiently developed to be used in priority setting. 
Consistent with the EDSTAC’s conclusions (USEPA, 1998a), the SAB/SAP
endorsed continued research to advance HTPS and bench-level in vitro
assay systems as well as (Q)SARs to provide measured or predicted data
concerning receptor binding and/or transcriptional activation that could
be used in future prioritization efforts.  The SAB/SAP also recommended
that, given the ambitious scope of the universe of chemicals the EPA
envisioned as potentially being included in the EDSP, the Agency
initiate Tier 1 screening with 50 to 100 chemicals. 

	Following review of public comment on an initial prioritization
proposal (USEPA, 2002), the Agency decided to select the first 50 to 100
chemicals based on exposure potential only (USEPA, 2005b).  The Agency
further determined that the first 50 to 100 chemicals selected for Tier
1 screening would be drawn from pesticide active ingredients as well as
pesticide inert ingredients with relatively large overall production
volumes when considering both pesticide and non-pesticide uses. In 2007,
the Agency published a final exposure-based prioritization scheme and a
proposed list of 73 pesticide chemicals to be subject to Tier 1
screening (USEPA, 2007a,b).  In 2009, the Agency published the final
list of 67 pesticide chemicals (58 active ingredients and 9 inert
ingredients) that will be subject to Tier 1 screening (USEPA, 2009c).  

	Consistent with EDSTAC and SAB/SAP recommendations the Agency continued
to develop in vitro (HTPS and bench-level platforms) and (Q)SAR methods
to support future prioritization efforts for chemicals in category 2
that contain little or no in vitro or in vivo effects data. The
Agency’s has on-going efforts to develop and evaluate Medium- and
High- Throughput assay methods (e.g., biochemical assays of protein
function, cell-based transcriptional reporter assays, multi-cell
interaction assays, transcriptomics on primary cell cultures, and
developmental assays in zebrafish embryos).  EPA will apply these
methods to over 300 food-use active ingredients, including ~50 of the 58
active ingredients in the initial group of 67 pesticide chemicals that
will be subject to Tier 1 screening. Several HTPS assays directly relate
to estrogen, androgen, thyroid and aromatase activity; results for the
pesticide active ingredients will be available at [ HYPERLINK
"http://www.epa.gov/comptox/toxcast" http://www.epa.gov/comptox/toxcast
]. Future HTPS efforts will include high production volume industrial
chemicals and pesticide inert ingredients.   

  	To advance prioritization techniques that include an effects
component, to complement an exposure-based prioritization method, the
Pesticide Program and EPA’s Office of Research and Development are
also developing (Q)SAR-based approaches to prioritize chemicals with an
initial emphasis on antimicrobial pesticides and inert ingredients used
in food-use pesticide formulations. The potential of these chemicals to
bind to the ER has been the initial focus of this effort. This paper
outlines the development of a (Q)SAR-based expert system and how it can
be used to prioritize food-use pesticide inert ingredients and
antimicrobial pesticides for potential estrogenicity through gaining an
understanding of a chemical’s potential to bind the ER.  

	The appropriateness of this expert system for prioritizing Tier 1
screening was evaluated using the OECD Principles for the Validation of
(Q)SARs (OECD, 2004a).  These principles are intended to guide the
(Q)SAR validation process undertaken within member countries and to
facilitate their regulatory acceptance by defining the types of
information that regulators would find useful when considering the
acceptability of individual (Q)SARs. These principles are also
consistent with, and complement, recommendations of the EDSTAC (USEPA,
1998a). The five OECD validation principles include: a defined endpoint;
a mechanistic interpretation; a defined domain of the (Q)SAR model
applicability; appropriate measures of goodness of fit, robustness,
ability to predict; and an unambiguous algorithm. Beyond the importance
of validating model equations and algorithms for national regulatory
needs, two important characteristics of (Q)SAR validations are addressed
by OECD to enhance regulatory acceptance for using estimated values for
priority-setting. The first characteristic is transparency of the (Q)SAR
estimate, not so much in terms of the methods used, but rather in terms
of how the estimate can be explained mechanistically and how the
estimate is reasonable based on data for comparable chemicals. The
second major characteristic for (Q)SAR acceptance is usefulness of a
particular (Q)SAR model for estimating endpoints of regulatory relevance
for all compounds within specified chemical inventories. Since the OECD
principles of validation for (Q)SAR seek to describe the domain of the
(Q)SAR model in terms of the chemical structures used to create the
model, usefulness can be evaluated by comparing the domain of the (Q)SAR
methods and the domain of the regulatory inventory assessed in a
specific regulatory context.  Unfortunately, most developers of (Q)SAR
models do not use a specific regulatory inventory as a target for the
model domain.  This document will discuss a (Q)SAR-based approach for
estrogenic activity that maintains mechanistic transparency and allows
the domain of the knowledge base to be aligned with the domains of
specific regulatory inventories. 

Evolution of the Expert System

The expert system reflects the culmination of research over the past
several years and input from scientists and regulators familiar with the
OECD (Q)SAR validation principles through two OECD forums. As part of a
Workshop on Structural Alerts for the OECD (Q)SAR Application Toolbox
(held on May15-16, 2008 in Utrecht, The Netherlands) an earlier
iteration of this work, a chemical sub-class approach to form a decision
support system for ER binding, was presented (OECD, 2009b) [see 
HYPERLINK "http://www.olis.oecd.org/olis/2009doc.nsf/linkto/"
http://www.olis.oecd.org/olis/2009doc.nsf/linkto/ env-jm-mono(2009)4].
The findings from this workshop indicated that the prototype system
could be useful for priority setting and appropriate for inclusion into
the OECD (Q)SAR Application Toolbox. The “Toolbox” is a stand-alone
software system developed by OECD for use by risk assessors around the
globe to facilitate the use of (Q)SAR systems documented in accordance
with OECD (Q)SAR validation principles (OECD, 2004a). Subsequent to this
workshop, an OECD Expert Consultation to Evaluate an Estrogen Receptor
Binding Affinity Model for Hazard Identification (OECD, 2009a) was held
February 16, 2009 in Paris, France. This was the first in a series of
OECD consultations to review expert systems and to expand the number of
toxicologically-based categories for use by risk assessors. The expert
system presented here reflects the peer input from the 2009 OECD
consultation. 

Chapter II. Mechanistic Basis of the Expert System to Predict Relative
Estrogen Receptor (ER) Binding Affinity

Overview of the Approach to Build an Effects-Based Expert System 

The conceptual approach providing the foundation for the expert system
is the description of the chemically-initiated perturbation of the ER
system in the form of an adverse outcome pathway. The ER-mediated
reproductive impairment adverse outcome pathway (Schmieder et al., 2004)
describes the linkage between the event that initiates the pathway
(e.g., a chemical binding the ER) and measures made at successively
higher and more complex levels of biological organization (Figure 1).
The pathway progresses from the molecular initiating event, through cell
and tissue level gene transcription and translation, continuing through
organ effects to an adverse outcome observed in the individual or a
population. With a plausible pathway to an adverse outcome described, a
rationale is provided for using the molecular initiating event as a
basis for prioritizing chemicals for further screening with EDSP Tier I
assays, which incorporate endpoints at higher levels of biological
organization.  The specific in vitro assays used to develop the expert
system are: i) measured chemical binding to the rainbow trout ER to
detect the potential for a chemical to initiate the ER-mediated pathway;
and ii) ER-mediated vitellogenin induction in rainbow trout liver slices
to confirm that ER binding translates to an effect at a point further
along the ER-mediated adverse outcome pathway (Schmieder et al., 2004). 

The four sections of Chapter II present the key components of the
approach to developing the expert system. The first section presents ER
binding affinity as an indicator of potential reproductive effects. The
next section reviews what was known of chemical interaction in the ER
binding domain. The third section described the regulatory chemical
domain of interest in this study and how the structures in the desired
inventories (i.e., food use pesticide inert ingredients and
antimicrobial active ingredients) were characterized. This section also
discusses the chemical domain that the expert system was developed to
cover and the extent to which existing information could be used in
developing the system. Finally, how the assays used to develop the
training set are optimized for measuring chemical-ER binding for
compounds in the chemical domain to ensure the collection of accurate
data is presented. 

Chapter III describes how understanding of the ER binding domain,
chemical domain, and assay domain in Chapter II is applied to establish
an ER binding training set representative of the inert and antimicrobial
chemical domains of interest. The chemicals in the training set make up
the expert system’s model domain, and thus they must adequately
represent the structures in the regulatory inventories. Chapter III also
describes how the knowledge gained from the training set can be
“coded” into systematic logic rules that establish a rule-based
expert system. 

Chapter IV of the report describes how the expert rules are implemented
in a decision tree to support prioritization of chemicals in the two
inventories for Tier 1 screening. Chapter IV also summarizes on-going
research on human ER binding of a strategically chosen subset of food
use inert ingredients and antimicrobial pesticides.  

ER Binding Affinity: An Indicator of Potential Reproductive Effects 

	There are multiple protein targets and pathways through which chemicals
may interfere with normal processes resulting in impaired reproduction. 
Predicting which chemicals may be capable of interfering with any given
pathway remains a challenge and many of the research efforts over the
past two decades in the area of endocrine disruption have focused on
increasing understanding of how chemicals perturb these endogenous
hormone systems. Despite the complexity of events that can lead to
reproductive impairment, it has long been appreciated that chemical
binding to the ER is one important conserved mechanism of interfering
with processes involved in reproduction. The ER-mediated adverse outcome
pathway (Figure 1) is a conceptual model that is useful to illustrate
how ER binding can be linked through a series of measurable events to
adverse outcomes of regulatory concern. As chemicals likely to initiate
or block the ER-mediated pathway are identified, the conceptual model is
useful for generating testable hypotheses at various levels of
biological organization along the pathway and provides risk assessors
and managers a readily articulated and testable rationale for
determining the order in which chemicals will undergo required screening
(i.e., prioritizing).	

	It is important to recognize that while the determination of potential
adverse reproductive effects in whole organisms or populations is an
important risk assessment issue, it is not a goal of a prioritization
application. The potency of a chemical for producing a specific adverse
outcome in vivo cannot be assumed to be the same as that for chemical-ER
binding in a cell-free assay due to many considerations including
chemical kinetics as well as interactions and feedback between cell and
organ systems.  While a broad generalization of the quantitative
relationship between ER binding affinity and in vivo response cannot be
drawn at this time, it is important to note that there is evidence that
chemicals with relatively low ER binding affinity can elicit significant
in vivo effects in fish, including reduced testicular growth, reduced
fecundity, changes in secondary sex character and behavior, testis-ova,
as well as complete phenotypic sex reversal (e.g., see Jobling et al.,
1996; Gray et al., 1999; Seki et al., 2003a,b; Yokota et al., 2005).
However, predicting in vivo potency of a chemical that can bind to the
ER is a more complicated and different problem than predicting relative
ER binding affinity to support an effects-based prioritization scheme
designed to focus in vivo testing that can determine whether or not
reproductive effects can occur, and if they do occur, at what dose. 

While the expert system reported here is designed to predict rainbow
trout ER binding affinity, based on a training set of binding data, an
in vitro rainbow trout tissue slice assay was employed to confirm
ER-mediated gene activation within a metabolically-competent tissue
(Gilroy et al., 1996).  Rainbow trout liver slices have been shown to be
ER responsive to parent chemicals and their metabolites (Shilling and
Williams, 2000; Schmieder et al., 2000; Schmieder et al., 2004). The
liver slice assay provides confirmation that the ER binding translates
to gene activation or antagonism at the next, higher level of biological
organization along the ER-mediated adverse outcome pathway (e.g.,
tissue/organ level; Figure 1) and increases confidence in the linkage
between low affinity ER binding (see final section of Chapter II),and
gene activation.  

The ER Binding Domain

	The energetic and steric constraints dictated by the ER itself, shapes
the domain of the chemical structures that can bind to the receptor.
Although only a small percentage of pesticide and industrial chemical
structures may be capable of binding, the ER is sufficiently promiscuous
to permit binding with a diverse array of chemical structures.  An
understanding of the energetic and steric characteristics of the ER
binding domain, therefore, provides the means to establish a mechanistic
basis for defining a chemical structure space associated with ER
ligands.

	Theories of chemical interaction in the various “sub-pockets”
within the ER ligand binding domain have been presented by numerous
investigators.  Theories for chemical interaction with the ER were
typically based on information gained from steroidal structures that
interact at two points within the ER i.e., two hydrogen-bonding groups
of a chemical within a specified distance interacting with receptor
protein in different ER sub-pockets.  However, it is also known that
chemicals that contain only one hydrogen-bonding group, such as
nonylphenols, bind ER and cause subsequent gene activation and in vivo
effects (Jobling et al., 1996; Seki et al., 2003b).  Katzenellenbogen et
al. (2003) and references there-in describe the “dynamic and plastic
character of ER” in their description of ER binding sub-pockets.
Through their larger body of work they have identified three primary ER
binding sub-pockets (referred to further in this paper as sites A, B,
and C; Figure 2) each with different requirements for hydrogen bonding.
This information proves useful for formulating and testing hypotheses of
how chemicals with only one potential hydrogen-bonding group may
interact with the ER, and how this interaction with the ER may vary
dependent upon the character of the hydrogen-bonding group. The focus of
this paper concerns chemical interactions at site A and B; site C is
most relevant to  investigations of potent synthetic ER antagonists (see
Katzenellenbogen et al., 2003 and references therein).

	Various (Q)SAR, docking, and conformational models also provide
insights as to how chemicals, primarily steroidal compounds or compounds
with relatively high ER binding affinity, may interact with the ER
(e.g., Weise and Brooks, 1994; Tong et al., 1997; Shi et al., 2002). A
review of (Q)SAR models developed largely from higher affinity ER
binding data is found in Schmieder et al., (2003a). These models,
however, do not appear adequate for interpreting interaction of
compounds with low ER binding affinities. In the approach described in
this paper, systematic measurements of chemicals with low binding
affinity were undertaken to establish a (Q)SAR training set that
reflected current understanding of the ER binding domain and was
representative of the chemical groups in the specific pesticide
inventories of interest.

The Regulatory Chemical Domain

	There were several aspects to consider in developing an expert system
to prioritize food use pesticide inert ingredients and antimicrobial
active ingredients for EDSP Tier I screening. First, the structures of
the chemicals found in the inventories needed to be characterized and
existing information examined to determine the extent to which ER
binding had previously been measured for the types of chemicals found on
the inventories. Where sufficient information was lacking, a systematic
approach (Schmieder et al., 2003b) to identify key chemicals for testing
was established to develop a training set representative of the diverse
chemicals in the two chemical inventories of interest. An iterative
approach included strategic selection of chemicals for measuring binding
affinity within chemical groups, assessment of results and trends within
chemical groups in the context of the ER binding hypotheses, and
additional selection of chemicals for measuring binding affinity until
sufficient information was gathered to make a prediction for all the
chemical groups in the inventories.  

Characterization of Inert and Antimicrobial Inventories. To establish
the chemical structures within the food use pesticide inert ingredients
(FI) and antimicrobial active ingredients (AM) inventories, the Chemical
Abstract System (CAS) number for each compound was obtained and then
cross-referenced to existing databases containing quality assured
chemical structures (see SOP for Chemical Characterization, Appendix I).
The FI inventory contained 893 total entries, of which 393 were discrete
chemical structures (93% organic, 6% inorganic, 1% organometallic). The
500 remaining entries in the inventory included polymers of mixed chain
length, chemical mixtures, and undefined substances (e.g., sand, blood,
tannins, cod liver oil).  The AM inventory contained a total of 299
substances of which 211 were discrete chemical structures (72% organic,
25% inorganic, 3% organometallic) and 88 were polymers, mixtures, and
undefined substances. The chemical sub-classes represented in these
inventories included chemicals without rings (i.e., “acyclic”
chemicals) such as alcohols, amines, quaternary amines, acyclic
sulfates, borates, carboxy esters, aldehydes, ketones, and phosphoric,
carboxylic, and sulfonic acids.  Chemical groups with ring structures
(i.e., “cyclic” chemicals) included alkylbenzenesulfonic acids,
alkylphenols, alkyloxyphenols, chlorobenzenes, hydrofurans, cyclic
hexanones and hexanols, imidazolidines, isothiazolines, oxazoles,
parabens, cyclic pentanones, phenones, pyrrolidines, phthalates,
salicylates, sorbitans, sulfonic acid dyes, triazines, and chemicals
with multiple functional groups further divided into “mixed phenols”
if a ring hydroxyl group was present, or “mixed organics” if there
was no a hydroxyl group. 

The Chemical Inventory Domains of Interest and Existing Information. The
early recognition that some chemicals interact with the ER and alter the
endocrine system has led to the development of large databases of ER
binding affinity.  The majority of early work in this area came from
drug design and the search for highly potent estrogen antagonists for
potential use in breast cancer therapies. Compounds in these databases
tend to be chemicals with binding affinities comparable to the strong
binding of the natural ligand and have structures similar in steric and
electronic characteristics.  Consequently, the diversity of structures
in these databases is limited with respect to defining a (Q)SAR domain
for chemicals not designed a priori to be biologically active (e.g.,
pesticidally inert ingredients) or with chemicals designed for a primary
biological activity not associated with endocrine effects (e.g., many
classes of pesticide active ingredients). Consequently, these existing
data are of limited use in developing (Q)SAR models to estimate ER
binding affinity for inventories of pesticide and industrial chemicals. 


While a previous review suggested that there was limited binding data
relevant for the inventories of interest (Schmieder et al. 2003a), to
ensure any existing binding affinity and/or transcription studies
relevant to the inventories of interest were incorporated in the current
effort, the literature was re-reviewed to determine if chemical groups
in the FI and AM inventories had been evaluated previously. The overlap
of FI and AM chemical structures with ER binding data reported in the
literature was found to be minimal. The chemicals with binding data
reported in the literature represented chemical classes that comprised
less than 4% of the chemicals in the FI and AM inventories.
Consequently, it was determined that a (Q)SAR model based on literature
data would not have a domain applicable to the regulatory inventories of
interest.

The Receptor Binding Assay Domain  

The historical focus on identifying chemicals with strong binding
affinities for potential drug development resulted in much of the early
ER binding affinity data to be generated using “cutoffs” in chemical
concentrations, with the result that many classes of chemicals with
lower binding potential were categorized as non-binding, even though
their binding potential was not fully characterized.  Consequently,
initial (Q)SAR models for ER binding that incorporated these data in
their training sets produced erroneous ER binding affinity predictions
for lower affinity chemicals when predicted values were compared to
empirical data derived from assays optimized for detection of low
affinity binding (Schmieder et al., 2004). 

	To address this issue, the research described here reflects an emphasis
on low affinity measurements for chemicals with a wide array of
physical-chemical properties, using ER binding and transactivation
assays optimized for detecting low affinity effects (for details see
Schmieder et al., 2004).  The resulting evaluation of binding affinity
at higher substrate concentrations than had been typically employed
required the development of methods to evaluate chemical solubility in
the assay medium and to interpret binding curves using additional
assays, e.g. Ki determinations or gene activation.  

Assays used to develop the training sets were designed to detect low,
but measurable, chemical-ER interactions and optimized for the types of
chemicals found in the specific regulatory inventories of interest. The
inventories of concern contain diverse structures with widely varying
physical-chemical properties (e.g., Log Kow of < -1.0 to > 8.0; neutral
organics; weak acids; organometallics, etc), thus chemical behavior in
the assays used in developing the binding data is an important
consideration. To limit false negative interpretation of measured
ER-binding (i.e., incorrectly assuming a chemical had no potential to
bind to the ER), chemicals were tested up to the limit of solubility in
the assay media employed. Thus, the maximum test concentrations is
determined on a chemical-by-chemical basis as chemical availability in
test media will vary with structure, aqueous solubility, Log Kow,
non-specific protein binding, pKa, etc. (see Heringa et al., 2004 for
protein effects on chemical availability in endocrine assays). To
illustrate the point, the maximum solubility of a chemical of Log Kow
1.3 in a protein enriched test system can be -1 Log M, while a chemical
of Log Kow 6 may be insoluble above -4 Log M. Chemical effects on the
assay components (e.g., protein denaturation) as well as assay effects
on chemical availability (e.g., pH affecting chemical speciation, total
protein content altering chemical availability) are also important to
consider to minimize misinterpretation of results, especially in
cell-free binding assays. To limit false positive interpretations (i.e.
incorrectly assuming a chemical is capable of binding to the ER),
representative chemicals within each structural class evaluated across
the associated Log Kow range, were tested in transactivation assays to
confirm ER-mediated gene activation. For chemical-ER interaction
determined in whole-cell or tissue assays, chemical-specific
cytotoxicity, or solubility in absence of cytotoxicity, were used to set
maximum test concentrations. These recommendations for testing chemicals
with a wide array of physical-chemical properties are also outlined in a
draft OECD Detailed Review Paper on Environmental Endocrine Disruptor
Screening: the Use of Receptor Binding and Transcriptional Activation
Assays for Fish (OECD, 2009c – draft in review). 

Chapter III.  Regulatory Inventories of Interest and Expert System Model
Domain

Overlaying Model Domains and Regulatory Inventories

The majority of (Q)SAR models reported in the literature (more than
15,000 models for a wide array of endpoints) are simple statistical
relationships between endpoint data and molecular descriptors (Bio-Loom,
2006).  The primary limitation of these statistical models is that
chemicals can produce the same endpoint activity even though the
toxicity mechanisms by which they interact with biological systems are
completely different.  For example, one chemical may bind to the ER via
a hydrogen bond donating group and another may bind via a hydrogen bond
accepting group or hydrophobic forces.  When data from chemicals that
elicit different mechanisms involving a common biological effect are
combined statistically in a single (Q)SAR model, the model is likely to
give spurious results unless the domain is narrowly defined (Bradbury et
al., 2003).  

	Consistent with the OECD principles for (Q)SAR validation, evaluation
of a (Q)SAR model for specific regulatory purposes should include an
evaluation of the mechanism of chemical-biological interaction (e.g.,
molecular initiating event in an adverse outcome pathway; Figure 1) in
addition to an evaluation of how well the endpoint value is predicted by
the model.  In general, statistical evaluations have not been successful
in characterizing the knowledge that discriminates interaction
mechanisms (Bradbury et al., 2003; OECD, 2009a). The OECD Guidance
Document on the Validation of (Quantitative) Structure-Activity
Relationship ((Q)SAR) Models (OECD, 2004a) offers several examples of
statistical clustering approaches for defining the domain of (Q)SAR
models using simplified structure spaces.  However, these methods should
be used with caution because the dimensionality of chemical structure
space is still not well defined.  

	As stated previously, a primary concern is how well the domain of a
(Q)SAR model represents the breadth of chemical structures included in
the regulatory inventory of interest (see Chapter II The Regulatory
Chemical Domain).  In the current effort, the regulatory applicability
issue is addressed by subdividing inert ingredients and antimicrobial
pesticides into respective subgroups that have specific chemical
attributes mechanistically related to ER binding.  As empirical data was
generated for chemicals within the subgroups, it was used to develop
“testable rules,” similar to those used for structural alerts and
other SAR techniques. The training sets (i.e., the knowledge base)
evolved as "expert rules” were tested with additional chemicals in the
inventory to determine if there was sufficient data to span the chemical
structures in specified subgroups. In this manner, the training set for
the expert system was systematically tested to ensure the knowledge base
represented the structure space of the two regulatory domains of
interest, i.e. food use inert ingredients and antimicrobial pesticides.

 Expanding the ER Binding Affinity Knowledge Base

The FI and AM knowledge base was established by selecting chemicals to
investigate binding mechanisms as well as chemicals to expand the
structural domain to overlap the regulatory inventories. Prototype
expert rules were established to address structural features or chemical
properties associated with binding types (see Chapter II. The ER Binding
Domain). Initial evaluation of these rules revealed that none of the FI
or AM chemicals contained two hydrogen-bonding groups at the distances
required for high affinity interactions (e.g., A-B binding of estradiol
and similar high affinity chemicals with hydrogen-bonding distances of
10.2 to 11 Å as described in Serifimova et al., 2007). Thus, no testing
was conducted for chemicals with these structural characteristics.
Instead, testing efforts were concentrated on better defining
chemical-ER binding within individual sub-pockets, specifically Site A
and Site B in the ER ligand binding domains.

Both inventories contain a significant number of chemicals with at least
one possible hydrogen bonding substituent. Due to the importance of
hydroxyl groups in hydrogen bond interactions in Site A, a series of
alkylphenols were tested for ER binding. While mostly para-substituted
alkylphenols were tested, a sufficient diversity of other phenols was
included to allow sub-grouping with respect to branched versus n-chain
and ortho, meta, and para substitution. A group of alkylphenols, from
the basic structure phenol with no substituents, to the 12 carbon chain
p-dodecylphenol, covering a Log Kow range from 1.5 to 8, were examined.
All chemicals within the series, were found to bind ER. Typical binding
displacement curves for this series of chemicals are shown in Figure 3A
along with that of the endogenous hormone 17ß-estradiol (E2). The EC50s
for each chemical is compared to that of the positive control (i.e.,
displacement of [3H]-E2 from the ER by unlabelled E2) to calculate the
relative binding affinity (RBA) for all chemicals relative to E2 at 100%
(see Schmieder et al., 2004 for details). 

 A subset of chemicals, representing the entire alkylphenol series and
the associated Log Kow extremes, were tested in trout liver tissue
slices to determine whether apparent ER binding resulted in gene
activation.  Transactivation in trout liver slices was demonstrated
across the Log Kow range of 1.5 to 8, using phenol and
p(n,t)-alkylphenols with C1 to C5, C8, C9, and C12 carbon chains (e.g.,
see Figure 3B for representative vitellogenin mRNA induction data for
phenol, 4-n-butylphenol, 4-t-octylphenol and 4-dodecylphenol). 

The potential for steric hindrance in ER binding for alkylphenols was
also investigated, with only the 2,6-ditertbutyl substituted compound
shown to be incapable of binding, presumably due to inability of the
phenolic hydroxyl group to get close enough to the A site to form a
hydrogen bond. Additionally, a group of p-alkoxy phenols was also
tested. All were determined to bind to the ER. 

The ER binding affinity of alkylphenols was found to increase linearly
with Log Kow from 1.5 to ~4.5 and then remain nearly constant with
increasing lipophilicity up to a Log Kow 8 (Figure 4).  The bilinear
relationship between ER binding affinity and Log Kow is consistent with
the hypothesis that binding can be increased through hydrophobic forces
in the center of the ER sub-pockets. Because Log Kow is readily
calculable from structure and has a mechanistic rationale, the RBA - Log
Kow relationship allows one to predict from the structure of an untested
chemical whether it belongs within the bounds of a tested chemical group
and, if so, whether it would be predicted to bind ER and produce
vitellogenin in a trout liver slice.  

Several chemical groups, which contain hydrogen bond acceptor
substituents, were also investigated for their potential to interact at
ER sub-pocket Site B. Initially, an extensive series of p-alkylanilines
were tested, analogous to the alkylphenols, to evaluate the potential
role of hydrogen bond interactions in receptor binding. The
p-alkylanilines from Log Kow of 1.3 to 5.2 were found to bind ER in a
relationship similar to that found with alkylphenols (Figure 4). Similar
p-alkoxyanilines also bound to the ER and followed the same relationship
with Log Kow (data not shown). The alkylanilines and the alkoxyanilines,
however, show less affinity for ER when compared to alkyl- and
alkoxyphenols of the same alkyl chain length (e.g., note ER RBA is
~0.001% for an alkylaniline of Log Kow 5 vs. ER RBA of ~0.01% for an
alkylphenol of Log Kow 5 in Figure 4). In this regard, aniline itself
was determined to not bind to the ER whereas phenol did bind and produce
gene activation. 

The p-alkylphenol and p-alkylaniline chemical series shown in Figure 4
represent “lead” groups for low affinity hydrogen-bond donors in A
Site, and low affinity hydrogen bond acceptors at B site. Additional
chemical groups likely to interact at Site A, as hydrogen-bond donors,
include parabens, salicylates with Log Kows of 1.8 to 5.5. Some of these
compounds have previously been identified in the literature as capable
of binding human ER and/or activating human ER mediated gene expression
(e.g., Routledge et al., 1998; Miller et al., 2001; Schultz et al.,
2000; Leskinen et al., 2005; Morohoshi et al., 2005; Terasaka et al.,
2006) and inducing ER-mediated production of vitellogenin in fish (e.g.,
Alslev et al., 2005; Bjerregaard et al., 2003). Although details are not
presented here, many additional chemical classes in the FI and AM
inventories that have hydrogen-bond acceptor substituents (e.g.,
phenones, phthalates, ring-substituted benzoates) were tested in the
current study and determined to have rainbow trout ER affinities similar
to those of alkylanilines with similar Log Kows. 

	Ultimately, a wide variety of chemicals were tested to represent all
the chemical subgroups in the FI and AM inventories. Chemicals in the
training sets were selected across the respective Log Kow ranges of each
subgroup in the inventories. In contrast to the groups previously
mentioned, these remaining chemical subgroups did not have compounds
capable of binding to the ER at the assay limit of detection. These
included both neutral and ionic organic chemicals, acyclic chemicals,
and organometallic compounds. The chemicals tested that showed no ER
binding affinity (RBA<0.00001%) are listed by group/subgroup in Table 1.
The results suggest that chemicals with no binding affinity do not
contain moieties with sufficient hydrogen bonding capability to interact
with ER or are excluded from the binding site because of hydrophilic
moieties. 

Some of the chemical groups that did not bind to the ER (i.e., RBA <
0.00001%) contained a large number of chemicals. As mentioned
previously, it is important to ensure the structure space within FI and
AM inventory chemical groups is represented in the training set
especially when a limited number of tested chemicals are representing a
much large chemical group. For instance, there were more than 70
alkylaromatic sulfonic acids in the inventories. Figure 5 shows
representative chemical structures that were tested from this group;
spanning a Log Kow range from -0.62 to 5.67. Another group, the sulfonic
acid dyes, were tested over the Log Kow range of -1.0 to 6.04 and found
not to bind ER (Figure 6). Figure 7 shows the same sulfonic acid dyes
presented in Figure 6 with some additional sulfonic acid dyes from the
inventory that were not tested (Figure 7 structures in white
background), but predicted to have an RBA less than 0.00001%, based on
being within the Log Kow of the knowledge base and thus within the
domain of dyes found to have measured RBA < 0.00001%. Figure 7 also
illustrates how chemicals in the training set can be compared to any
untested inventory chemical thereby allowing users of the expert system
to evaluate the adequacy of the model for their prioritization
application.

Another important characteristic of the alkylaromatic sulfonic acids,
and the sulfonic acid dyes is that these chemicals are charged in the
assay solutions. A chemical carrying a charge is less likely to
partition across membranes or into hydrophobic pockets. It is also
possible that charged chemicals interact with and denature receptor
protein. Chemicals with a sulfonic acid group, found in both
alkylaromatic sulfonic acids and sulfonic acid dyes, possess an
estimated pKa of 0.70, which would suggest complete ionization to
charged species in the pH 7.4 buffer media. Interestingly, while most
chemicals in these charged groups (Figure 5 and 6) did not give an
indication of binding, occasionally chemicals with a net ionic charge
gave apparent binding displacement (see OBS and DBS in Figure 5A- ER
binding displacement graphs). However, follow-up experiments to detect
the nature of the binding interaction (see Ligand Binding
Characterization Studies and kaempferol data in Schmieder et al., 2004
for details) and/or assays using trout slices designed to detect results
of ER agonism or antagonism (Figure 5B) at the tissue level, showed no
evidence of ER-mediated gene activation or interruption of estradiol
driven vitellogenin induction, respectively. As previously mentioned, it
is highly likely that ‘apparent’ binding of charged compounds that
do not translate into gene activation, or ‘apparent’ binding that is
subsequently shown to be non-competitive at the ER ligand binding
domain, is instead due to charged species denaturing the ER protein in
binding assays and causing release of [3H]-estradiol (i.e.,
non-competitive, endogenous ligand displacement). Because this
‘apparent’ binding can result in release of the ligands used in
displacement assays and yield false positives (e.g., OBS and DBS in
Figure 5A), it is recommended that apparent ER binding for charged
chemicals be confirmed using higher order assays. Additional chemical
groups possessing a charge in assay media include those with carboxylic
acid moieties (with an estimated pKa of 4.1 making them completely
ionized at pH 7.4), and quaternary amines with a permanent charge.
Representatives of these groups tested in the assays were all negative
(RBA < 0.00001%; Table 1). 

In addition to chemical charge, another important consideration is the
potential for chemical hydrolysis in the assay solution, resulting in
formation of a chemical species in the assay system. For example, the
borate esters did not bind ER (RBA < 0.00001%) and this inactivity may
be explained by the formation of an acyclic species with a Log Kow less
than that needed for interaction with ER (i.e., these acids likely
hydrolyze resulting in the formation of boric acid). 

	Some chemical groups not found in the inventories were tested to
increase confidence in understanding structural attributes that preclude
the possibility of ER binding. Negative binding results (RBA < 0.00001%)
and lack of agonist or antagonist activities in trout slices are shown
in Figure 8 (A & B) for non-ring substituted benzoates with Log Kow
values of 3.59 to 4.76.  This finding is in contrast to ER binding
activity (RBA > 0.00001%) found for similar benzoate structures that
have additional substituents at position 2, or position 4, or position
2, 3, and 4 on the benzene ring. 

It is beyond the scope of this summary to recount all the subgroups of
chemicals tested. Nonetheless, this iterative process to establish the
FI and AM training sets was continued until all the respective groups
were examined. There are chemicals on the inventories that could not be
easily assigned to a single subgroup. These chemicals have multiple
functional groups; e.g., “mixed phenols” if they also contain at
least one hydroxyl moiety making them capable of hydrogen-bond donation
and interaction at site A; or “mixed organics” if they do not
contain a hydroxyl-ring substituent but contain more than one
substituent that might be a hydrogen-bond acceptor. A total of 16 mixed
phenols were tested in the ER binding assay. Of these, three showed some
affinity for the ER (e.g., RBA > 0.00001%) and 13 had no affinity for ER
(RBA < 0.00001%).  Additionally, there were a total of 47 mixed organics
tested, with five showing binding (RBA > 0.00001%).  As with other
groups, some of the chemicals in the mixed phenols and mixed organics
training sets are not in the FI or AM inventories but were tested to
investigate hypothesized binding mechanisms. Structures of mixed phenols
and mixed organics in the current training set are shown in Appendix II
and III, respectively, to illustrate the diversity of these chemicals.
While the majority of mixed phenols and mixed organics that have been
tested thus far do not bind to the trout ER, there is not sufficient
data at this time to further categorize these compounds into subgroups
which bind ER and those which do not, or to hypothesize structural
parameters associated with the activity of those compounds. 

There are additional groups for which a limited number of chemicals were
tested; e.g., substituted benzoates, gallates, and thiophosphate esters.
While the current  training set contains sufficient knowledge to make
predictions for these chemicals within the respective inventories, there
is insufficient knowledge to make predictions for chemicals within these
groups that are outside the chemical domain on the FI and AM
inventories.  If a chemical is classified in one of these subgroups and
it is not an exact match to a training set structure the chemicals is
determined to be outside of the model domain and an “unknown binding
potential” outcome is assigned. As additional chemical inventories are
evaluated and the training set expanded to represent a larger number and
greater diversity of mixed phenols, mixed organics, and other groups
under represented subgroups, it should be possible to gain a better
understanding of substituent patterns that result in ER binding and
provide the means to add additional rules to the expert system. 

A Rule-Based Expert System to Predict ER Binding Affinity

	Chemical structures in heterogeneous inventories are inherently
hierarchical.  When some chemical groups behave quite differently from
others, or the information known about each group varies substantially,
rule-based expert systems offer one approach to estimating specific
properties of an entire array of chemicals.  Creating decision trees
that reflect the hierarchy and applying expert rules and localized
(Q)SAR models for specific chemical structures has numerous advantages
in a prioritization scheme.  Perhaps most important is the advantage of
maintaining transparency of the estimates by describing the logic for
grouping the chemicals, providing the localized (Q)SAR model or expert
rules that were used to make the estimate, providing empirically
measured data for chemicals in the associated training set, and
providing literature values and citations for other chemicals in the
same group to support the estimates.  These data enable users of the
expert system to discuss the scientific basis for the all estimates, and
justify modifying a specific estimate if new knowledge becomes available
to the user.  

	Figure 9 presents an overview of the expert system’s decision tree
for ER binding affinity.  The current version examines each chemical in
the respective inventories (see Figure 9, “Start Here”) and places
them into groups of inactive chemicals (if the answer to query I is
‘No’,  query II is ‘Yes’, or query III is ‘Yes’), into
“drug-like” groups of chemicals that have the potential for high
affinity ER binding (‘Yes’ to query IV), or into groups of chemicals
that may have weak-to-moderate binding affinity depending on specific
properties or structural features and answers to queries V, VI, and VII.
 Each of these major queries is sub-divided as needed to more explicitly
describe chemical subgroups associated with a binding type and
properties of the training set chemicals associated with RBA > 0.00001%
or < 0.00001% . Examples of questions that are used to categorize
chemicals beyond what is shown in summary Figure 9 are provided in
Figures 10, 11, and 12, which are expansions of queries I, V, and VII,
with explanatory text provided below. 

	Regarding query I, it has been generally accepted that chemicals must
have at least one cycle (e.g., saturated or aromatic carbon ring) to
competitively bind at the ER, although previously, there was limited
evidence to fully evaluate this assumption.  This general “testable
rule” is mechanistically reasonable because flexibility and binding
affinity, in general, vary inversely (Katzenellenbogen et al., 2003 and
references therein; Serafimova et al., 2007).  Cyclic and multi-cyclic
chemicals tend to be less flexible, at least in the absence of long
alkyl chains. Given the large percentage of acyclic chemicals in the
inventories of interest, numerous acyclic chemicals were tested and none
had measureable binding (i.e., none with RBA > 0.00001%). These included
both ionic (quaternary amines, sulfates, and phosphates) and non-ionic
acyclic structures (e.g., alcohols, amines, esters; see Figure 10).
Specific acyclic chemicals tested are listed in Table 1. Therefore,
query I (see Figure 9) in the expert system asks whether or not the
chemical contains a cycle.  If it does not contain a cycle, the chemical
is not predicted to bind to the trout ER.   

	Query II in Figure 9 asks if a cyclic chemical contains a charge. If it
does, and it belongs to one of the groups listed in Figure 9 and within
the specified Log Kow ranges the chemical is predicted to have an RBA <
0.00001%. If the answer to query II in Figure 9 is ‘No’, then the
decision tree invokes the next query. 

An uncharged cyclic chemical moving on to query III in Figure 9, is
examined to determine if it belongs to any one of the subgroups of
chemicals for which members within the Log Kow ranges tested thus far
had RBA < 0.00001%.  There are a variety of steric and electrostatic
reasons why these groups cannot bind to the ER.  Generally, the absence
of hydrogen bonding groups, or inappropriate geometry explains the
failure of these chemicals to bind to ER (Katzenellenbogen et al., 2003
and references therein). Suffice it to say, these inactive groups are
well-defined and the training sets showed no members of these groups
have had an RBA > 0.00001%.  The “known inactive” subgroups include:
alkyl benzthiols, p,n-alkyl fluorobenzenes, a hindered alkylphenol,
benzamides, borate esters, benzoates lacking ring substituents,
bis-anilines, hydrofurans (alcohol and ketone), imidazolidines,
isothiazolines, mono-cyclic hydrocarbons, oxazoles, n-alkyl phenones,
pyrrolidiones, sorbitans, and triazines (Figure 9; Table 1). Note that
the range of Log Kow covered by the tested chemicals within each
specific chemical group is shown in Figure 9.  A chemical within the
group, but outside the tested Log Kow range is not within the domain of
the expert system and no binding potential is assigned. 

	The chemicals that do not meet the requirements of the first three
queries belong to groups of chemicals for which some members of their
respective subgroup bind ER.  Chemicals within these subgroups with the
greatest binding affinity to the ER receptor are those that mimic
steroids and have two oxygen atoms available for hydrogen bonding at the
specific distances dictated by the ER receptor.  Chemicals that meet
these requirements are removed from the other subgroups for further
evaluation (Query IV); note that none of the chemicals currently in the
FI or AM inventories have structural characteristics defined in this
query.  Chemicals that do not meet the requirements for high affinity
binding are further evaluated by the expert system (Figure 9; note arrow
going from ‘No’ in query IV and on to query V). 

To evaluate these remaining subgroups, chemicals are next queried
against a set of “special rules”.  Query V was established for
chemicals with unique structural features in the training set (see
Figure 11; note that none of the FI or AM chemicals satisfy rules for
DDT-like, tamoxifen-like, and multicyclic hydrocarbons groups) and/or to
ensure coverage of FI and AM inventory chemicals (i.e.,
p-alkylchlorobenzene and thiophosphate ester groups). Currently training
sets for these subgroups are sufficient to represent the FI and AM
inventories; however, the expert system will not make a prediction for
any chemicals in these subgroups that are outside the model domain. 
However, users of the expert system could over ride the system if they
determined structures in the training set are sufficiently similar to a
compound of interest and can be used as analogs to assign a RBA
classification. 

	If a chemical does not meet any of the requirements for the first five
queries, it is evaluated further to determine if it has structural
features that make it likely to interact at site A or B. Initially the
chemical’s Kow is estimated. We have noted that of over 350 chemicals
tested in the training set, no chemical with a Log Kow less than 1.3 has
been shown to competitively bind with ER. Thus, chemicals with a Log Kow
less than 1.3 are assigned a RBA of less than 0.00001%. While low
affinity binding does not appear to occur for any chemical with Log Kow
< 1.3, the reverse cannot be said; i.e., it is not the case that all
chemicals with Log Kow > 1.3 will bind ER. There are more specific (more
restrictive) Log Kow ranges for binding activity that apply within many
of the subgroups. These ranges are listed in Figure 9 for each
group/subgroup as applicable. 

	The sixth query identifies chemicals with Log Kow > 1.3 that have a
phenolic substituent as a hydrogen bond donating group; these chemicals
are presumed to have the potential to bind at Site A in the ER (VI in
Figure 9). This includes the p-alkylphenol series previously described,
p-alkoxyphenols, parabens, and salicylates. Specific Log Kow ranges for
active chemicals (RBA > 0.00001%) in each subgroup are listed. Details
of each training set are not included here but ER binding and trout
liver slice transactivation data comparable to that shown for the
alkylphenols in Figure 3A & 3B, as well as depictions of structures of
each training set vs. the inventory structure (as in Figure 7) are
available for all the groups listed in Figure 9.  It is this latter
information that is especially important to the evaluation of mixed
phenols. While the presence of the phenolic group makes it a possible
hydrogen-bond donor, with potential binding at Site A, there are
additional substituents present that could hinder binding activity.  As
described previously, currently there is insufficient information to
establish general rules for these types of chemicals. A chemical falling
in this group is first evaluated for an exact match to an existing mixed
phenol in the training set chemical (Appendix II) and is assigned the
appropriate RBA. Only 3 of the 13 mixed phenols tested have been found
to bind ER above the RBA detection limit of 0.00001%.  However, if a
chemical is not an exact match to a tested mixed phenol, it is
considered to have an ‘unknown binding potential’.  

	Query VII in Figure 9 identifies chemicals with aniline, ester, or
other polar substituents that can serve as a hydrogen bond acceptor at
the B site in the ER (Figure 2).  The expanded query is shown in Figure
12. The knowledge base of chemicals currently thought to bind at this
site contains: p-alkylanilines, p-alkoxyanilines, phthalates, branched
phenones, p-alkyl cyclohexanols and hexanones, and various
ring-substituted benzoates.  Again, each one of these groups has
specific Log Kow ranges where ER binding has been measured. Similar to
the mixed phenols in query VI, query VII for ‘mixed organics’
compares a compound to the mixed organic training set (Appendix III); if
there is an exact structural match the corresponding RBA values are
assigned.  A mixed organic chemical that is not in the training set is
characterized as having an “unknown binding potential” (i.e., the
chemical is characterized as outside the model domain); however, as
discussed above the user can override the system if they determine a
chemical(s) in the training set can serve as an analog.   

	Using the rules summarized in Figure 9, the expert system was applied
to the FI and AM inventories. The predictions obtained from the system
are summarized in Table 2. For the FI inventory, of 393 discrete
chemicals, 378 are predicted to have a RBA less than 0.00001% (the limit
of detection in the assay) and 15 are predicted to have an RBA of
>0.00001%. For the AM inventory, of 211 discrete chemicals, 196 are
predicted to have an RBA less than 0.00001% and 15 are predicted to have
an RBA greater than 0.00001%. Note that some of the structures predicted
to have RBA < 0.00001% fell into groups that had some chemicals that
bind ER, but the specific structure in question did not have a high
enough Log Kow to bind ER.  For example, 4-cyclohexanols of Log Kow 2.4
to 3.8 in the training set were found to bind ER, but cyclohexanols with
Kows below this cutoff did not bind to the ER.  

	As mentioned previously, if the RBA for any chemical cannot be
predicted, the chemical contains structural features which are not
included in the chemicals in the current training set and are not
currently in the model domain.  The approach, however, is based on a
perspective that the domain of the expert system knowledge base can be
expanded with strategic testing of additional chemicals in inventories
of regulatory interest.  This hypothesis based, strategic expansion of
the expert system’s domain provides a systematic basis to incorporate
additional chemical diversity.  

	

Chapter IV.  Accessibility of the Expert System

Expert Systems for ER Binding Affinity

	One of the more straight forward approaches to developing an expert
system for a complex endpoint is to use decision trees with the queries
based on simple “if, then else” logic.  In current chemical decision
trees, the conditions placed on individual queries in order for them to
be logically true can range from simple presence or absence of a general
substructure to a complex parameter range derived from still other
molecular models.  As illustrated in this work, a decision tree model
has been developed that incorporates simple chemical substructure
queries, i.e., expert rules for when to invoke (Q)SAR models for ER
binding affinity. 

The reason expert system technology is so important is that SARs are
typically most reliable for chemicals that interact through the same
mechanisms.   This underlying SAR principle would mean that the user
would have to be familiar with the multidimensional interactions of
chemicals with biological systems. Obviously, this complex reality
places an undue burden on users of the models.  The solution is to
encode the expert knowledge of chemical interactions and provide the
user with a fully transparent description of what considerations were
made, how the estimate was calculated from these considerations, and
literature citation(s) to underpin the expert rules. 

	The use of decision trees facilitates regulatory acceptance because
they inform the user about how well the decision tree model covers the
domain of the chemicals to be assessed (usefulness) and, at the same
time, provide a transparent explanation for the basis of the estimated
endpoint.  Explaining the “why” behind each chemical prediction in
the decision tree is crucial to subsequent discussions in a
prioritization application. For example, since binding to a receptor has
many specific constraints that preclude structures from fitting within
the receptor “pocket(s),” a large number of chemical groups are
expected to be “nonbinding.” To build a decision tree to estimate ER
binding affinity, a series of queries to identify nonbinding groups of
chemicals was established and combined with a series of more specific
rules to group large categories of chemicals with a range of binding
affinities. Each of these categories was further classified and
quantitative models to estimate binding affinity for the entire category
was included. Even more specific structural requirements were added to
identify steroid-like chemicals of which many had large binding
affinities similar to drugs and potent estrogenic chemicals.  

The expert system for ER binding affinity was created with specific
chemical inventories and domains. The expert system can be easily
expanded as additional data become available on chemicals from
additional domains. The expert system model was also developed with the
presumption that the model could be used manually for individual
chemicals, or totally automated on a small personal computer for ranking
larger lists of chemicals. The query language for the logical grouping
of chemicals as well as the estimation of ER binding affinity is written
in a syntax that is understandable and transferable to computer
languages. As discussed in Chapter I, upon completion of the expert
system, consistent with SAP review comments, EPA envisions including the
expert system in OECD’s (Q)SAR toolbox.

Ongoing Research: Human ER Binding Affinity and Gene Activation

ERα.  Specifically, a subset of the FI and AM chemicals are being
tested with a binding assay using full-length recombinant human   SEQ
CHAPTER \h \r 1 ERα and a transactivation assay using human T47D-KBLuc
cells. Chemicals are selected for human ER in vitro testing based on
predictions from the current expert system to provide coverage of each
chemical subgroup and Log Kow range within the FI and AM inventories. 
Preliminary data suggests that the same groups of chemicals bind ER from
both species, although additional subgroups remain to be evaluated.
Within the subgroups evaluated thus far, the trend is that fewer members
of a chemical group bind to the human ER than rainbow trout ER (e.g., a
more restrictive Log Kow range for binding human ER within a group).    

These findings support previous studies that noted a concordance across
species (fish, rat, human) when similar assay systems are compared
(i.e., chemical affinity for recombinant ER from species A vs.
recombinant ER from species B, or chemical affinity for cytosolic ER
from species A vs. cytosolic ER from species B, or comparing affinities
measured in assays with similar total protein concentrations and
chemical availability). Binding affinities measured using recombinant
ERs (generally with low total assay protein) are typically orders of
magnitude greater than affinities measured using ER derived from
cytosolic preparations where total assay protein can be 10 times
greater. This is likely due to differences in non-specific binding of
chemicals to proteins (Schmieder and Henry, 1988) other than estrogen
receptor protein (see Heringa et al., 2004 for more discussion of the
influence of assay total protein on measured ER activity).  For
instance, RBAs for select alkylphenols and monohydroxy parabens tested
using ER from rat uterine cytosol (Blair et al., 2000) are within same
order of magnitude of affinities determined in this study and in
published literature for rainbow trout ER when receptors are isolated
from liver cytosol (Olsen et al., 2005).  Consistent with the
relationships noted above,  ER binding affinities for these same
chemicals, determined using recombinant ERs from either human or trout
origin, are often an order of magnitude greater than those determined
using cytosolic receptors. Other researchers have also shown that there
is good ER binding concordance across species for lower affinity
chemicals (Matthews et al., 2000; Petit et al., 1997; Olsen et al.,
2005). However care must be taken when comparing studies if cutoffs in
maximum concentration tested differ between studies, which will confound
the data interpretation. 

The comparative work on human ER for FI and AM chemicals will continue,
using the systematic approach described, to characterize any
differences, and if needed. tailor expert systems that are specific to
‘a vertebrate’ (i.e., fish or human) or to establish an expert 
system that will be protective of the most sensitive species for a
specific chemical subgroup. 

V. Conclusions 

	The EDSTAC (USEPA, 1998a) and SAB/SAP (USEPA, 1999) outlined the
strengths and limitations of using (Q)SARs as an effects-based component
in future prioritization schemes for chemicals that have  little or no
empirical effects data.  The strengths of using (Q)SAR systems include
the means to rapidly predict biological activities of large numbers of
chemicals, at little or no cost, and thereby avoid the need to
prioritize on the basis of “no data.”  The EDSTAC also concluded
that (Q)SARs can provide the means to prioritize chemicals in a
transparent and consistent manner that avoids the problem of comparing
data derived from different experimental systems.  The EDSTAC noted that
(Q)SARs that predict receptor binding and/or gene transcription
activation, for example, have the same strengths and limitations as the
experimental data derived from in vitro assays.  In that regard, the
EDSTAC cautioned that use of experimental or predicted receptor binding
or gene activation data does not necessarily predict in vivo responses,
nor does it encompass other mechanisms of endocrine disruption.  As
guiding principles for development and use of (Q)SARs for priority
setting the EDSTAC also recommended that applicable chemical domains be
sufficiently diverse for their intended application, the training sets
be developed using the most complete and accurate data sets available,
and that (Q)SARs be validated and used only within the range of
conditions for which they were developed and evaluated.  

The OECD Principles for the Validation of (Q)SARs (OECD, 2004a) provide
a detailed framework to address the perspectives developed by the
EDSTAC.  The OECD principles include documentation of a well-defined
biological endpoint; a mechanistic basis for the model; a defined domain
of model applicability; appropriate measures of goodness of fit,
robustness and predictability; and an unambiguous algorithm. 
Transparency is a critical characteristic of the OECD (Q)SAR validation
principles – particularly in the context of the mechanistic
plausibility of the predictions and the reasonableness of estimated
values compared to experimentally derived data;  the chemical domain of
a (Q)SAR application; and the regulatory purpose of the (Q)SAR.     	

The ER binding expert system described here represents an encapsulation
of both undirected research published in the literature from the past
two decades as well as directed research over the past four years to
tailor (Q)SAR methods to provide effects information to support future
development of prioritization schemes for Tier 1 EDSP screening.
Development of the systems was guided by the OECD principles of (Q)SAR
validation and the (Q)SAR-based expert system reported here is derived
from a training set of food use inert ingredients and antimicrobial
active ingredient pesticides.  When combined with estimates of exposure
potential, consistent with EDSTAC (USEPA, 1998a) and SAB/SAP (USEPA,
1999) recommendations, the expert system provides a complementary
effect-based means to prioritize food use inert ingredients and
antimicrobial pesticides for future screening and testing.   Given that
this (Q)SAR system is specific to ER binding, Tier 1 prioritization
would be relevant for those Tier 1 assays that are specific to, or
capable of, detecting potential interactions with the estrogen system;
e.g., uterine cytosolic ER binding assay, ER transactivation assay,
uterotrophic (rat) assay, pubertal female (rat) assay, and the fish
short term reproduction assay
(http://www.epa.gov/scipoly/oscpendo/pubs/fish_fs.pdf).    

The expert system for ER binding reported here can be readily expanded
to address additional chemical inventories.  Prioritization of chemicals
for Tier 1 assays designed to detect potential interactions with the
androgen of thyroid hormone systems would require corresponding expert
systems specific to those receptors.  While the current expert system is
specific to ER binding affinity, the underlying approach for developing
the system could be applied to the development of training sets for
other nuclear receptors to further expand hazard-based prioritization
schemes. 

 	Research and development on HTPS and bench-level in vitro and (Q)SAR
methods had not sufficiently progressed for use in the initial round of 
EDSP priority setting. Contingent on the SAP’s review, and completion
of the corresponding human ER binding and gene activation data sets to
confirm the preliminary findings of concordance between rainbow trout
and human data, the (Q)SAR-expert system would provide effects
information relevant to the estrogen hormone system for use in
subsequent rounds of EDSP priority setting for food use inert
ingredients and antimicrobial pesticides  An option for future Tier 1
screening of chemicals identified through a (Q)SAR-informed
prioritization scheme for chemicals with little or no existing in vivo
data could include a phased approach in which in vitro Tier 1 assay data
is collected and evaluated to inform prioritization decisions on the in
vivo Tier 1 screening assays.

	In conclusion, the Agency is exploring the expansion of the existing
expert system to non-food use inert ingredients, which include many of
the same chemical classes represented by the food use inert ingredient
and antimicrobial inventories. In an iterative process, bench-level or
HTPS data will be used to develop and validate an evolving (Q)SAR-based
expert system for predicting estrogen-based effects data. The same
principles of training set and (Q)SAR development could also be
systematically applied to specified inventories to address other
receptor-based mechanisms of endocrine disruption; e.g. the androgen
hormone system.  Decisions regarding future efforts will include an
evaluation of forthcoming HTPS data and the extent to which current HTPS
and/or bench level assays methods will require any modification to
support development of appropriate (Q)SAR training sets for chemicals
with low receptor binding affinity. Ongoing research measuring in vivo
responses to chemicals with low ER binding affinity will also inform
interpretation of predicted or empirically measured in vitro data in
EDSP prioritization and screening.

References

Alslev, B., B. Korsgaard, and P. Bjerregaard. 2005. “Estrogenicity of
Butylparaben in Rainbow Trout Oncorhynchus mykiss Exposed via Food and
Water.” Aquatic Toxicology.  72: 295-304.

Bio-Loom, “Ver. 1.5 Using CLogP.” 2006. Biobyte Corporation,
Claremont, CA, USA.. http://www.biobyte.com

Bjerregaard, P., Andersen, D. N., Pedersen, K. L., Pedersen, S. N., and
Korsgaard, B. 2003. “Estrogenic Effect of Propylparaben
(propylhydroxybenzoate) in Rainbow Trout Oncorhynchus mykiss After
Exposure via Food and Water.” Computational Biochemistry and
Physiology. Part C. 136: 309-317.

Blair, R. M., Fang, H., Branham, W., Hass, B., Dial, S., Moland, C.,
Tong, W., Shi, L., Perkins, R., and Sheehan, D. 2000. “The Estrogen
Receptor Relative Binding Affinities of 188 Natural and Xenochemicals:
Structural Diversity of Ligands.” Toxicological Sciences. 54: 138-153.

Bradbury S. P., Russom, C.L., Ankley, G.T., Schultz, T. W., and Walker,
J.D. 2003. “Overview of Data and Conceptual Approaches for Derivation
of Quantitative Structure-Activity Relationships for Ecotoxicological
Effects of Organic Chemicals” Environmental Toxicology and Chemistry.
22(8): 1789-1798.

Bradbury S. P., Feijtel, T. C., and Van Leeuwen, C. J. 2004. “Meeting
the Scientific Needs of Ecological Risk Assessment in a Regulatory
Context.” Environmental Science and Technology. 38: 463A-470A

Gilroy, D. J., Miranda, C. L., Siddens, L. K., Zhang, Q., Buhler, D. R.,
and Curtis, L. R. 1996. “Dieldrin Pretreatment Alters [14C]dieldrin
and [3H]7,12-dimethylbenz[a]anthracene Uptake in Rainbow Trout Liver
Slices.”  Fundamental and Applied Toxicology. 30: 187-193.

Gray, M. A., Niimi, A. J., and Metcalfe, C. D. 1999. “Factors
Affecting the Development of Testis-ova in Medaka, Oryzias latipes,
Exposed to Octylphenol.” Environmental Toxicology and Chemistry. 18:
1835–1842.

Heringa, M., Schreurs, R., Busser, F., Van Der Saag, P., Van Der Burg,
B., and Hermens, J. 2004. “Toward More Useful In Vitro Toxicity Data
With Measured Free Concentrations.” Environmental Science & Technology
38: 6263–6270.

Jobling, S., Sheahan, D., Osborne, J. A., Matthiessen, P., and Sumpter,
J. P. 1996. “Inhibition of Testicular Growth in Rainbow Trout
(Oncorhynchus mykiss) Exposed to Estrogenic Alkylphenolic Chemicals.”
Environmental Toxicology and Chemistry. 15: 194-202.

Katzenellenbogen, J., Muthyala, R., and Katzenellenbogen, B. S. 2003.
“Nature of the Ligand-binding Pocket of Estrogen Receptor α and
ß: The Search for Subtype-selective Ligands and Implications for the
Prediction of Estrogenic Activity.”  Pure and Applied Chemistry. 75:
2397-2403.

Leskinen, P., Michelini, D., Picard, D., Karp, M., and Virta, M. 2005.
“Bioluminescent Yeast Assays for Detecting Estrogenic and Androgenic
Activity in Different Matrices.” Chemosphere. 61: 259-266.

Matthews, J., Celius, T., Halgren, R., and Zacharewski, T. 2000.
“Differential Estrogen Receptor Binding of Estrogenic Substances: a
Species Comparison.”  Journal of Steroid Biochemistry and Molecular
Biology. 74: 223-234.

Miller, D., Wheals, B. B., Beresford, N., and Sumpter, J. P. 2001.
“Estrogenic Activity of Phenolic Additives Determined by an In Vitro
Yeast Bioassay.” Environmental Health Perspectives. 109(2): 133-138. 

Morohoshi, K., Yamamoto, H., Kamata, R., Shiraishi, F., Koda, T., and
Morita, M. 2005. “Estrogenic Activity of 37 Components of Commercial
Sunscreen Lotions Evaluated by In Vitro Assays. Toxicology In Vitro. 19:
457-469.

National Research Council (NRC). 2007. Toxicity Testing in the 21st
Century: A Vision and a Strategy. National Academy of Sciences,
Washington, D.C. http://www.nap.edu/catalog.php?record_id=11970#toc

Olsen, C., Meussen-Elholm, E., Hongslo, J., Stenersen, J., and
Tollefsen, K. 2005. "Estrogenic Effects of Environmental Chemicals: An
Interspecies Comparison.” Computational Biochemistry and Physiology.
Part C. 141: 267-274.

Organization for Economic Cooperation and Development, (OECD), 2004a.
The Report From The Expert Group On (Quantitative) Structure-Activity
Relationships [(Q)SARs] On The Principles For The Validation Of (Q)SARs.
Series on Testing and Assessment. No. 49. ENV/JM/MONO(2004)24.
17-Dec-2004.   HYPERLINK
"http://www.oecd.org/dataoecd/33/37/37849783.pdf" 
http://www.oecd.org/dataoecd/33/37/37849783.pdf 

OECD, 2004b. Chemical Testing Guidelines for Endocrine Disruptors:
Conceptual Framework. Jan. 2004.
http://www.oecd.org/document/58/0,3343,en_2649_34377_2348794_1_1_1_1,00.
html

OECD. 2008. Workshop on Integrated Approaches to Testing and Assessment.
Series on Testing and Assessment No. 88. ENV/JM/MONO(2008(10).
27-May-2008. http://www.oecd.org/dataoecd/45/52/40705314.pdf

OECD, 2009a. Report of the Expert Consultation to Evaluate an Estrogen
Receptor Binding Affinity Model for Hazard Identification. Task Force on
Hazard Assessment. ENV/JM/HA/RD(2009)1. 2-March-2009 107pp.  (to be
declassified 31-Jul-2009).

OECD. 2009b. Report Of The Workshop On Structural Alerts For The OECD
(Q)SAR Application Toolbox. Series on Testing and Assessment No. 101.
Feb., 2009. ENV/JM/MONO(2009)4.
http://www.olis.oecd.org/olis/2009doc.nsf/linkto/env-jm-mono(2009)4

OECD, 2009c. Draft Detailed Review Paper on Environmental Endocrine
Disruptor Screening: The Use of Receptor Binding and Transcriptional
Activation Assays for Fish.  Series on Testing and Assessment. No. XX.
May. 2009.
http://www.oecd.org/document/12/0,3343,en_2649_34377_1898188_1_1_1_1,00.
html  (public comment period ended 29-May-2009)

Petit, F., LeGoff, P., Cravedi, J. P., Valotaire, Y., and Pakdel, F.
1997.  "Two Complementary Bioassays for Screening the Estrogenic
Potency of Xenobiotics:  Recombinant Yeast for Trout Estrogen Receptor
and Trout Hepatocyte Cultures.” Journal of Molecular Endocrinology.
19: 321-335.

Routledge, E. J., Parker, J., Odum, J., Ashby, J., and Sumpter, J. 1998.
“Some Alkyl Hydroxy Benzoate Preservatives (Parabens) Are
Estrogenic.” Toxicology and Applied Pharmacology. 153: 12-19.

Schmieder, P.K. and Henry, T.R. 1988. “Plasma Binding of n-Butanol,
Phenol, Nitrobenzene, and Pentachlorophenol in Rainbow Trout and Rat: a
Comparative Study.” Computational Biochemistry and Physiology. 91C(2):
413-418.

Schmieder, P., Tapper, M., Linnum, A., Denny, J., Kolanczyk, R., and
Johnson, R. 2000. “Optimization of a Precision-cut Trout Liver Tissue
Slice Assay as a Screen for Vitellogenin Induction: Comparison of Slice
Incubation Techniques.” Aquatic Toxicology. 49: 251-268.

Schmieder, P., Ankley, G., Mekenyan, O., Walker, J., and Bradbury, S.
2003a. “QSAR Models for Predicting ER Binding Affinity of Structurally
Diverse Chemicals.” Environmental Toxicology and Chemistry. 22(8):
1844-1854. 

Schmieder, P., Mekenyan, O., Bradbury, S., and Veith, G. 2003b. “QSAR
Prioritization of Chemical Inventories for Endocrine Disruptor
Testing.” Pure and Applied Chemistry. 75 (11/12): 2389-2396.

Schmieder, P. K., Tapper, M. A., Denny, J. S., Kolanczyk, R. C., Sheedy,
B. R., Henry, T. R., and Veith, G. D. 2004. “Use of Trout Liver Slices
to Enhance Mechanistic Interpretation of Estrogen Receptor Binding for
Cost-effective Prioritization of Chemicals within Large Inventories.”
Environmental Science and Technology. 38(23): 6333-6342.

Schultz, T.W., Sinks, G., and Cronin, M. 2000. “Effects of Substituent
Size and Dimensionality on Potency of Phenolic Xenoestrognes Evaluated
with a Recombinant Yeast Assay.” Environmental Toxicology and
Chemistry. 19: 2637-2642.

Serafimova, R., Todorov, M., Nedelcheva, D., Pavlov, T., Akahori, Y.,
Nakai, M., and Mekenyan, O. 2007. “QSAR and Mechanistic Interpretation
of Estrogen Receptor Binding”. SAR and QSAR in Environmental Research.
18: 389-421.

Seki, M., Yokota, H., Matsubara, H., Maeda, M., Tadokoro, H., and
Kobayashi, K. 2003a. “Fish Full Life Cycle Testing for the Weak
Estrogen 4-t-pentylphenol on Medaka (Oryzias latipes).” Environmental
Toxicology and Chemistry. 22: 1487-1496.

Seki, M., Yokota, H., Maeda, M., Tadokoro, H., Kobayashi, K. 2003b.
“Effect of 4-Nonylphenol and 4-tert-Octylphenol on the Sex
Differentiation in Medaka (Oryzias latipes).” Environmental Toxicology
and Chemistry. 22: 1507–1516.

Shi, L., Tong, W., Fang, H., Perkins, R., Wu, J., Tu, M., Blair, R.,
Branham, W., Waller, C., Walker, J., and Sheehan, D. M. 2002. “An
Integrated ‘4-Phase’ Approach for Setting Endocrine Disruption
Screening Priorities - Phase I and II Predictions of Estrogen Receptor
Binding Affinity.” SAR and QSAR in Environmental Research. 13(1):
69-88.

Shilling, A. and Williams, D. 2000. “Determining Relative
Estrogenicity by Quantifying Vitellogenin Induction in Rainbow Trout
Liver Slices.” Toxicology and Applied Pharmacology. 164: 330-335.

Terasaka, S., Inoue, A., Tanju, M., and Kiyama, R. 2006. “Expression
Profiling of Estrogen-Responsive Genes in Breast Cancer Cells Treated
with Alkylphenols, Chlorinated Phenols, Parabens, or Bis- and
Benzoylphenols for Evaluation of Estrogenic Activity.” Toxicological
Letters. 163: 130-141.

	Tong, W., Perkins, R., Strelitz, R., Collantes, E., Keenan, S., Welsh,
W., Branham, W., and Sheehan D. M. 1997. “Quantitative-Structure
Activity Relationships (QSARs) for Estrogen Binding to the Estrogen
Receptor: Prediction Across Species.”  Environmental Health
Perspectives.  105: 1116-1124.

United States Environmental Protection Agency (USEPA). 1998a. Endocrine
Disruptor Screening and Testing Advisory Committee (EDSTAC) Final
Report: Chapter 4: Priority Setting. Aug. 1998.
http://www.epa.gov/scipoly/oscpendo/pubs/edspoverview/finalrpt.htm

USEPA. 1998b. EDSP: FRN: Proposed Statement of Policy. Vol. 63, No. 248.
Dec. 1998. 71542-71568. http://www.epa.gov/endo/pubs/122898frnotice.pdf

USEPA. 1999. Review of the EPA’s Proposed Environmental EDSP. July
1999.
http://yosemite.epa.gov/sab/sabproduct.nsf/C8ABD410E357DBCF85257193004C4
2C4/$File/ec13.pdf

USEPA. 2002. EDSP: FRN: Proposed Chemical Selection Approach for Initial
Screening. Vol. 67, No. 250. Dec. 2002. 79611-79629.
http://www.epa.gov/scipoly/oscpendo/pubs/12-02-frnotice.pdf

USEPA. 2005a. OncoLogic User’s Manual: An Expert System for Prediction
of the Carcinogenic Potential of Chemicals. Ver. 6.0. March 2005. 
http://www.epa.gov/oppt/sf/pubs/onco-user-man.pdf

USEPA. 2005b. Endocrine Disruptor Screening Program (EDSP): Federal
Registry Notice (FRN): Chemical Selection Approach for Initial
Screening. Vol. 70, No. 186. Sep. 2005. 56449-56465.
http://www.epa.gov/fedrgstr/EPA-TOX/2005/September/Day-27/t19260.pdf

USEPA. 2007a. EDSP: FRN: Draft List of Initial Pesticide Active
Ingredients and Pesticide Inerts to be Considered for Screening under
the Federal Food, Drug and Cosmetic Act. Vol. 72, No. 116. June 2007.
33486-33503. http://www.epa.gov/endo/pubs/draft_list_frn_061807.pdf

USEPA. 2007b. EDSP: FRN: Draft Policies and Procedures. Vol. 72, No.
239. Dec. 2007. 70842-70862.
http://www.epa.gov/scipoly/oscpendo/pubs/draft_policies_frn.pdf

USEPA. 2008a. EDSP: FRN: Data Requirements for Antimicrobial Pesticides.
Part III. 40 CFR Parts 158 and 161. Vol. 73, No. 196. Oct. 2008.
http://edocket.access.gpo.gov/2008/E8-23127.htm

USEPA. 2008b. Use of Structure-Activity Relationship (SAR) Information
and Quantitative SAR (QSAR) Modeling for Fulfilling Data Requirements
for Antimicrobial Pesticide Chemicals and Informing EPA’s Risk
Management Process. EPA-HQ-OPP-2008-0110-0045.
http://www.regulations.gov/fdmspublic/component/main?main=DocketDetail&d
=EPA-HQ-OPP-2008-0110

USEPA. 2009a. ASsessment Tools for the Evaluation of Risk (ASTER).
http://www.epa.gov/med/Prods_Pubs/aster.htm

USEPA. 2009b. Ecological Structure Activity Relationships (ECOSAR). Feb.
2009. http://www.epa.gov/oppt/newchems/tools/21ecosar.htm

USEPA. 2009c. EDSP: FRN: Final List of Initial Pesticide Active
Ingredients and Pesticide Inert Ingredients to be Screened under the
Federal Food, Drug, and Cosmetic Act. Vol. 74, No. 71. April 2009.
17579-17585. http://www.epa.gov/endo/pubs/final_list_frn_041509.pdf

Wiese, T. E., and Brooks, S. C. 1994. “Molecular Modeling of Steroidal
Estrogens:  Novel Conformations and Their Role in Biological
Activity.” Journal of Steroid Biochemistry Molecular Biology. 50:
61-73.

Yokota, H., Abe, T., Nakai, M., Murakami, H., Eto, C., and Yakabe, Y.
2005. “Effects of 4-tert-Pentylphenol on the Gene Expression of P450
11ß-Hydroxylase in the Gonad of Medaka (Oryzias latipes).” Aquatic
Toxicology. 71: 121-132. 

Figure 1. An estrogen receptor (ER)-mediated reproductive impairment
adverse outcome pathway  (from Schmieder et al., 2004) The pathway
demonstrates typical responses that might be observed at various levels
of biological organization and the types of assays (in vitro or in vivo)
from which observations along the pathway can be made. In this example,
chemical binding to the rainbow trout ER, the point of pathway
initiation, is used to develop a quantitative structure-activity
relationship (in silico model). 



Figure 2. The estrogen receptor ligand binding domain showing
sub-pockets A, B, C where chemical ligand hydrogen bonding interactions
occur.  Receptor protein amino acids involved in interactions with
chemical ligands are indicated by letter and sequence number. 
17ß-Estradiol is shown in the binding pocket (after Katzenellenbogen,
et al., 2003 and references therein, and pers. comm.).

Figure 3. Binding and gene expression of p-alkylphenols in rainbow
trout estrogen receptor (rtER) in vitro assays. A) rtER binding to
receptor derived from liver cytosol determined by displacement of
[3H]-17ß-estradiol (E2) for: unlabelled 17ß-estradiol (E2,(), phenol
((), 4-ethylphenol ((), 4-n-butylphenol ((), 4-n-octylphenol ((),
4-t-octylphenol ((), 4-n-nonylphenol ((), and 4-t-nonylphenol isomer
mixture ((); and, B) rtER mediated gene expression in trout liver
slices; symbols are matched for positive control 17ß-estradiol (left
side of each figure) and phenol or substituted p-alkylphenol (right side
of each figure) measured in the same experiment;  phenol ((, left Y-axis
scale); 4-n-butylphenol ((, right Y-axis scale); 4-tert-octylphenol ((,
middle right Y-axis scale); and 4-dodecylphenol ((, far right Y-axis
scale).

Figure 4.  Relationship between rainbow trout estrogen receptor binding
(Relative Binding Affinity; RBA) and Log Kow for chemicals interacting
at Site A (e.g., p-alkylphenols; blue diamond) or Site B (e.g.,
p,n-alkylanilines; green triangle).

Figure 5. Structures of alkylaromatic sulfonic acids, abbreviations,
and Log Kow (measured or estimated) and results of rainbow trout
estrogen receptor (rtER) binding assays illustrating: (A) absence of
rtER binding for PTS ((), EBS ((), 2NSA ((), NSA (() from -7 logM to
solubility for each, and apparent binding for OBS (() and DBS ((),
compared to 17ß-estradiol (E2) (-10 to -6 logM, (); and confirmed lack
of trout liver slice vitellogenin (VTG) mRNA induction for OBS and DBS
in: (B) agonist assay with E2 (-10 to -7 logM, symbols match chemical
exposures on the same day) or 6 concentrations of DBS ((, left Y-axis
scale) or OBS ((, right Y-axis, -5.25 to -3 logM); and (B) antagonist
assay with a maximum induction concentration of E2 (-8 logM) combined
with each of 5 concentrations of DBS (-5.3 to -3.5  logM,(, left Y-axis
scale), or E2 (-7 logM) combined with each of 5 concentrations of OBS
(-5 to -3 logM, (, far right Y-axis scale).

 

Figure 6. Structures of sulfonic acid dyes, abbreviations, and Log Kow,
and results of rainbow trout estrogen receptor (rtER) assays
illustrating absence of : (A) rtER binding for 9 dyes AYE ((); SYE (();
RAC ((); MIT ((); ABL ((); FBR ((); BCA ((); BAD (();WRB ((), compared
to 17ß-estradiol (E2) ((); and, (B) trout liver slice vitellogenin
(VTG) mRNA induction for 2 dyes compared to E2 positive control in
agonist assays WRB (() and AYE ((); and, B) antagonist assay showing no
change in the maximum VTG mRNA in response to -8 log M (() E2 when
combined with AYE from -5 logM to -3 logM.

 

 

 

Figure 7.   To provide transparency at each query in the expert system
existing data in the knowledge-base (i.e., data in the training set) is
compared to food use inert ingredient and antimicrobial pesticide
inventory chemicals. The example shows the sulfonic acid dyes structures
in the knowledge base (grey shaded structures at top of figure) and nine
representative chemicals from the food use inert ingredient inventory;
the first inventory structure shown is also on the antimicrobial
inventory. Chemical structures with no shading have not been assayed for
ER binding. Chemicals are presented in increasing order of Log Kow.
Since all dyes shown are within the Log Kow of the knowledge base, they
are in the domain and predicted to have RBA < 0.00001%.  

Figure 8. Structures of non-ring substituted benzoates, Log Kow and
their abbreviations and results of rainbow trout estrogen receptor
(rtER) binding assays illustrating: (A) absence of rtER binding for BB
((), IAB ((), NHE ((), PB ((), and PEB (() from -6 logM to solubility
for each benzoate compared to 17ß-estradiol (E2) (-10 to -6 logM, ();
and, (B) lack of trout liver slice vitellogenin (VTG) mRNA induction for
IAB in agonist assay with E2 (-10 to -5 logM, () or 6 concentrations of
IAB (-4.5 to -2.5  logM, (, left Y-axis scale); and, B) antagonist assay
with a maximum induction concentration of E2 (-8logM (()) combined with
each of 6 concentrations of IAB from -4.5 to -2.5  logM (() compared to
positive control E2 (-10 to -5 logM, () response slices run in the same
experiment. 

Figure 9.  The expert system decision tree based on rainbow trout
estrogen receptor binding affinity and trout liver slice transactivation
knowledgebase.  

Figure 10. Expansion of Decision Tree Query I.  When the answer to the
question “Contains a Cycle” is ‘No’, the chemical is determined
to be acyclic. The system then asks if the chemical contains a charge.
Whether or not the chemical contains a charge, the same conclusion of an
RBA < 0.00001%  is reached.  This figure lists the specific chemical
groups in the training set from which this conclusion is based. No
rainbow trout estrogen receptor binding was detected for any of the 25
acyclic compounds tested across all chemical groups listed. Note that
acyclic compounds with “mixed functional groups” tested contained
multiple substituents from many of the groups listed below as well as
carboxylic acid and sulfonic acid substituents. Detection limit of the
assay under test conditions in this study is 0.00001% relative to
estradiol at 100% (RBA < 0.00001%).  

Figure 11. Expansion of Decision Tree Query V – Special Structure
Rules. Chemical queries are used to assign chemicals to groups/subgroups
witin Query V. A chemical satisfying the general group query is then
assessed for an exact chemical match to a training set chemical
structure and assigned the binding affinity of that tested chemical. If
an exact match is not found for the following groups, then “unknown
binding potential is assigned: DDT-type, multicyclic hydrocarbons,
tamoxifen-type and thiophosphate esters. However, the 4-alkyl
halobenzenes were sufficiently tested to establish sub-groups based upon
Log Kow cutoffs. For the 4-alkylchlorobenzenes, the Log Kow cutoff is
4.5, above which binding was observed (RBA > 0.00001%) for Log Kow 4.5
to 5.5.  No 4-alkylchlorobenzenes of Log Kow > 5.5 was tested, so an
“unknown binding potential” is assigned for chemicals in this group
with a Log Kow > 5.5. Two additional groups, alkylbromobenzenes and
alkyliodobenzenes have not been tested and are currently out of the
domain of this expert system. Detection limit of the assay under these
conditions is 0.00001% relative to estradiol at 100% (RBA < 0.00001%).  

Figure 12. Expansion of Decision Tree Query VII – Site B- Belongs to
known Active Class. A chemical not satisfying any previous query is
queried for structural fragments that would place it in one of the
chemical groups listed in the figure. Once assigned to a group its Log
Kow is compared to the specific Log Kow range of group members with an
RBA > or < 0.00001% assigned. If the Log Kow falls outside of specified
ranges the chemical is considered to have an “unknown binding
potential”. If a chemical contains more than one specific fragment
associated with B-site binding it is considered a “mixed organic”
chemical. The chemical is then assessed for an exact match to a training
set structure. If there is an exact match the associated, measured RBA
is assigned. If the chemical does not match a chemical in the training
set, the chemical is out of the expert system domain and an “unknown
binding potential” is assigned.

Table 1. Training set chemicals within chemical sub-groups for which no
rainbow trout estrogen receptor binding was observed at an RBA detection
limit of 0.00001%. 

Acyclic Chemicals	CAS	Chemical Name	LogKow

Acyclic borates	121437	Trimethylborate	-1.00

Acyclic alcohol	77996	1,1,1-tris(hydroxymethyl)propane	-1.00

Acyclic alcohol	111273	1-hexanol	2.03

Acyclic alcohol	104767	2-ethyl-1-hexanol	3.01

Acyclic alcohol	112301	1-decanol	4.24

Acyclic amine	112572	1,2-ethanediamine,
n-(2-aminoethyl)-n'-2-(2-aminoethyl)aminoethyl	-1.00

Acyclic amine	78900	1,2-diaminopropane	-0.91

Acyclic amine	2783177	1,12-diaminedodecane	3.51

Acyclic carboxy ester	97632	Ethyl methacrylate	1.94

Acyclic carboxy ester	689894	Methylsorbate	1.96

Acyclic aldehyde	4313035	(2E,_4E)-2,4-heptadienal	1.86

Acyclic ketone	504201	2,6-dimethyl-2,5-heptadien-4-one	2.68

Acyclic mixed	67436	Diethylenetriamine pentaacetic acid	-1.00

Acyclic mixed	139899	Trisodium (2-hydroxyethyl)ethylenediaminetriacetic
acid	-1.00

Acyclic mixed	55406536	3-iodo-2-propynyl N-butylcarbamate	2.65

Acyclic mixed	38916426	DL-Aspartic acid,
N-(3-carboxy-1-oxo-3-sulfopropyl)-N-octadecyl-, tetrasodium salt	5.33

Acyclic mixed	67633630
N-ethyl-N,N-dimethyl-3-[(1-oxoisooctadecyl)amino]-1-propanaminium, ethyl
sulfate	6.43

Acyclic mixed	637398	triethanolamine hydrochloride	-1.00

Acyclic mixed	42808366	Octadecanoic acid, 9(or_10)-(sulfooxy)-,
1-butylester, sodium salt	6.76

Acyclic phosphoric acid	4672382	propylphosphonic acid	0.28

Acyclic quat.amine	35141367	1-Propanaminium,
N,N,N,-trimethyl-3(trimethoxysilyl)-, chloride; 1-propanaminium,	-1.00

Acyclic quat.amine	56375792	tributylmethylammonium chloride	0.24

Acyclic quat.amine	112005	N-dodecyltrimethylammonium chloride	1.22

Acyclic sulfate	126921	sodium 2-ethylhexylsulfate	2.75

Acyclic sulfate	142314	sodium octylsulfate	2.88





	Cyclic  groups/subgroups



	Alkylaromatic Sulfonic Acids	657841	sodium_p-toluenesulfonate	-0.62

Alkylaromatic Sulfonic Acids	98691	4-ethylbenzenesulfonic_acid	0.38

Alkylaromatic Sulfonic Acids	532025	naphthalene-2-sulfonic_acid_sodium
salt	0.63

Alkylaromatic Sulfonic Acids	130143	Naphthalene-1-sulfonic acid, sodium
salt	0.92

Alkylaromatic Sulfonic Acids	6149037	4-Octylbenzenesulfonic acid, sodium
salt	3.56

Alkylaromatic Sulfonic Acids	27176870	Dodecylbenzenesulfonic_acid_	5.67





	Alkyl Benzthiols	106456	p-toluene_thiol	2.88

Alkyl Benzthiols	2396681	4-tert-butylthiophenol	4.19





	p-Alkylfluorobenzene	462066	Fluorobenzene	2.27

p-Alkylfluorobenzene	403394	1-fluoro-4-isopropylbenzene	3.71

p-Alkylfluorobenzene	28593148	4-fluoropentylbenzene	4.90





	Benzamide	10546700	N-propylbenzamide	1.72

Benzamide	--	2-methyl-n-propyl-benzamide	2.08





	Benzoates - non-ring subst	93992	Phenylbenzoate	3.59

Benzoates - non-ring subst	136607	n-butylbenzoate	3.84

Benzoates - non-ring subst	6789884	Hexyl benzoate; Benzoic acid, n-hexyl
ester	3.84

Benzoates - non-ring subst	94473	Phenethylbenzoate	4.01

Benzoates - non-ring subst	94462	Isoamylbenzoate	4.15





	Bis-Anilines	101779	4,4'-diaminodiphenylmethane	1.59

Bis-Anilines	54628216	4,4'-diamino-2,2'-dimethylbibenyl	3.57

Bis-Anilines	4073987	4,4'-methylenebis(2,6-dimethylaniline)	4.01

Bis-Anilines	101688	1,1'-methylenebis 4-isocyanato benzene	4.08

Bis-Anilines	101611	4,4'-methylenebis[N,N-dimethyl]aniline	4.45

Bis-Anilines	2390592
Ethanaminium,N-[4-[bis[4-(diethylamino)phenyl]methylene]-2,5-cyclohexadi
en-1-ylidene]-N-ethyl-,chloride	5.78

Bis-Anilines	2716101
4,4'-[1,4-phenylenebis(1-methylethylidene)]bis-benzenamine	6.04

Bis-Anilines	13680358	4,4'-methylenebis(2,6-diethylanilne)	6.15





	Cyclic Alcohols - other	96413	Cyclopentanol	0.71

Cyclic Ketones - other	120923	Cyclopentanone	0.38

Cyclic Ketones - other	78591	Isophorone	1.70

Cyclic Ketones - other	4694126	2,4,4-trimethylcyclopentanone	1.86

Cyclic Ketones - other	76222	(+/-)-Camphor	2.18





	Furans- sugar	57501	Sucrose	-1.00

hydroFurans	97994	tetrahydrofurfuryl_alcohol	0.02

hydroFurans	105215	5-butyldihydrofuran-2(3H)-one	1.15





	Imidazolidines	116256	dimethyl-1-hydroxymethylhydantoin	-0.86

Imidazolidines	16079882	1-bromo-3-chloro-5,5-dimethyl hydantoin	0.39

Imidazolidines	77485	1,3-dibromo-5,5-dimethyl-2,4-imidazolidinedione
0.54

Imidazolidines	118525	1,3-dichloro-5,5-dimethylhydantoin	0.55





	Isothiazolines	2634335	1,2-benzisothiazol-3-one	0.61

Isothiazolines	26530201	2-octyl-3(2H)-isothiazolone	2.45





	Mono-Cyclic Hydrocarbons	108883	Toluene	2.73

Mono-Cyclic Hydrocarbons	106423	p-Xylene (anhydrous)	3.15

Mono-Cyclic Hydrocarbons	98828	Cumene	3.66

Mono-Cyclic Hydrocarbons	5989275	(R)-(+)-Limonene	4.57



 

	Oxazoles	6542376	1-aza-3,7-dioxabicyclo(3.3.0) octane-5-methanol	-1.00

Oxazoles	7747355	5-ethyl-1-aza-3,7-dioxabicyclo(3.3.0)-octane	0.26





	Phenones - n-chain	98862	Acetophenone	1.63

Phenones - n-chain	942927	n-hexanophenone	3.64

Phenones - n-chain	1671756	Heptanophenone	4.13





	Pyrrolidiones	872504	1-methyl-2-pyrrolidone	-0.38

Pyrrolidiones	2687947	N-octylpyrrolidone; 1-Octyl-2-pyrrolidone	3.31





	Sorbitans	1338392	Sorbitan,_monododecanoate_(9CI)	3.15

Sorbitans	1338438	(z)-sorbitan,_mono-9-octadecenoate	5.89





	Sulfonic Acid Dyes	1934210	Acid_Yellow_23	-1.00

Sulfonic Acid Dyes	2783940	C.I._Food_Yellow_3	-0.44

Sulfonic Acid Dyes	25956176	C.I. Food Red 17	0.08

Sulfonic Acid Dyes	3567257	Sulcofuron-sodium, monohydrate	1.89

Sulfonic Acid Dyes	3844459	C.I. Acid Blue 9, disodium salt	3.03

Sulfonic Acid Dyes	4404437	Fluorescent Brightener 28	3.09

Sulfonic Acid Dyes	5281049	Pigment Red 57-1	3.41

Sulfonic Acid Dyes	27344418	Disodium 4,4'-bis(2-sulfostyryl)biphenyl
3.84

Sulfonic Acid Dyes	7023612
2-Naphthalenecarboxylicacid,4-[(5-chloro-4-methyl-2-sulfophenyl)azo]-3-h
ydroxy-,calcium salt(1:1)	6.04





	Triazines	7673098	N,N',N-trichloro-1,3,5-triazine-2,4,6-triamine	-0.38

Triazines	2893789	Dichloro-s-triazinetrione	1.28

Triazines	5915413
6-chloro-n(1,1-dimethylethyl)-n'-ethyl-1,3,5-triazine-2,4-diamine	3.06

Triazines	28159980
N'-tert-butyl-n-cyclopropyl-6-(methylthio)-1,3,5-triazine-2,4-diamine
3.38



Table 2.  Number of discrete chemicals in the food use inert ingredient
(FI) and antimicrobial pesticide (AM) inventories that are predicted to
have rainbow trout estrogen receptor binding affinities below or above
an RBA of 0.00001% based on the expert system depicted in Figure 9.. 

  

											FI	AM

Number of chemicals in Acyclic, Charged, or Known Inactive Groups
predicted to have RBAs < 0.00001%

Acyclics (all groups)									230	122

Alkylaromatic Sulfonic Acids	  			  				  78	    3

Sulfonic Acid Dyes		   			    				    9	    1

Sorbitans			   			    				    7	    0

Monocyclic Hydrocarbons     				    				    7            0

Cyclic caged Hydrocarbons	   			    				    1	    0

Pyrrolidiones			 	 		    				    3	    0

Hydrofurans		    	 	 		    				    3	    0

n-Alkyl Phenones		   	 		    				    	    1	    0

Oxazoles			    	 		   		 		    0	    3

Triazines			    	 		   		 		    1	  13

Isothiazolines		    	 	 				    		    3	    7

Imidazolidines	 	    	 	 				    		    0	    8

Cyclic Inorganics		    	 				    		    	    0	    3

2,6-subst Alkylphenol		    	 				    		    1	    0

Cyclic Pentanones/Others	    	 				    		    	    2  	    0

Number of Chemicals in Site A, B, or Special Rule Groups predicted to
have RBAs < 0.00001% because of Log Kow specific cut-offs

4-Cyclohexanones of Log Kow <2.4	 		   		 		    1             0

4-Cyclohexanols of Log Kow <2.4	    	 	   		 		    	    1             0

4-Chlorobenzenes of Log Kow <4	    	 	   		 		  	    1   	     0

Number of Mixed Functional/Heteratom Chemicals with Log Kow < 1.3 and/or
measured 

RBA < 0.00001% 

Mixed Phenols		    	 		   					    7	    3

Mixed Organics		   			  					  20	  29

Organometallics		     	 		    					    2	    4

Total Discrete Chemicals with RBAs predicted to be <0.00001%		378	196

							

Chemicals predicted to have RBAs > 0.00001% by satisfying Site A, B, or
Special Rule queries and Log Kow > group-specific cut-off 						

Alkyl Phenols   	   	   								   3	    9 

Alkoxy Phenols		   								   1   	    0

Parabens		    	   	 						   3  	    0

Salicylates	    	   	   							   1    	    0

Phthalates                          	   								   3    	    0

Thiophosphate Esters	   	   							   1    	    0

Mixed Functional/Heteratom Chemicals with Log Kow >1.3 and measured RBA
> 0.00001% 

Mixed Phenols		   								   2    	    6 

Mixed Organics		   								   1    	    0 

Total Discrete Chemicals with RBAs predicted to be >0.00001%		  15    	 
15

________________________________________________________________________
___

Total Discrete Chemicals in the Inventory					393	 211							 

 PAGE   

 PAGE   vi 

 PAGE   44 

  PAGE   \* MERGEFORMAT  61 

 PAGE   

 PAGE   63 

B)

 

A)

 

B)

A)

Log Kow = 0.38 calc

 

Log Kow = 0.63 msrd

 

2NSA

Log Kow = 0.92 calc

 

NSA

Log Kow = -0.62 msrd

 

PTS

Log Kow = 3.56 calc

 

OBS

Log Kow = 5.67 calc

 

DBS

EBS

Log Kow = -1.00 calc

 

AYE

Log Kow =  -0.44 calc

 

SYE

Log Kow = 0.08 clog

 

RAC

Log Kow = 1.89 msrd

 

MIT

Log Kow = 3.03 clog

 

ABL

Log Kow = 3.09 clog

 

FBR

Log Kow = 3.41 clog

 

BCA

Log Kow = 3.84 clog

 

BAD

Log Kow = 6.04 avg

 

WRB

 

A)

B)

%

2

T

U

=

h8

h8

 h8

 h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

  h8

hó

hó

h`

h8

hÃ

hÃ

hÃ

hHp

h

h

h

h

h

hG

hG

h

h8

_

h‰

h

t

š

	

!

,

/

¿

Ã

h¡I

h¡I

h

h

  h8

 h8

  h8

  h8

 h8

 h8

 h8

 h8

  h8

h8

h

/ h8

) h8

  h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

  h8

 h8

ᔠh8

  h8

  h8

  h8

  h8

) h8

, h8

 h8

  h8

 h8

 h8

 h8

  h8

  h8

 h8

 h8

 h8

>* h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

ùñùêßÔùÍùÍùÅ½Åùñù¶ùÅù¯ù¶ùÅù¶ù§ùÅŸÅù¶
ù§ùÅùñù¶Å•ÅùŠ  h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

  h8

 h8

 h8

 h8

ࠆᔁ㡨ᰋᘀ뽨唀Ĉ  h8

 h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

  h8

  h8

 h8

  h8

 h8

 h8

 h8

 h8

öîãÜÔÜÍÜ¾¯ ‘‚{Ü{s{l{l{a{ZO{H h8

  h8

 h8

hŒ

  h8

 h8

h\(

 h8

>* h8

 h8

 h8

 h8

 h8

hŒ

 h8

 h8

 h8

>* h8

  h8

 h8

hŒ

 h8

 h8

 h8

  h8

  h8

  h8

# h8

# h8

  h8

  h8

 h8

 h8

>* h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

  h8

 h8

>* h8

 h8

 h8

 h8

 h8

 h8

 h8

>*  h8

>* h8

 h8

  h8

  h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

# h8

 h8

 h8

 h8

 h8

  h8

  h8

 h8

 h8

 h8

 h8

 h8

 h8

  h8

( h8

& h8

  h8

# h8

+ h8

( h8

  h8

# h8

# h8

) h8

  h8

  h8

+ h8

& h8

# h8

# h8

# h8

  h8

# h8

( h8

1 h8

& h8

# h8

# h8

# h8

  h8

  h8

) h8

, h8

4  h8

# h8

& h8

# h8

# h8

# h8

( h8

# h8

  h8

# h8

( h8

 h8

 h8

 h8

 h8

 h8

 h8

4 h8

) h8

, h8

# h8

# h8

) h8

) h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

 h8

h

$ h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

h8

 

BB

Log Kow = 3.84 msrd

 

IAB

Log Kow = 4.15 msrd

 

NHE

Log Kow = 4.76 est

 

PB

Log Kow = 3.59 msrd

 

PEB

Log Kow = 4.01 msrd

 

A)

B)

