Predictive Models

The EDKB program has developed computer-based models to predict affinity for binding of compounds to the estrogen and androgen nuclear receptor proteins. The process to develop the models was based on a close linkage of the laboratory and the modeling that resulted in training sets appropriately designed to calibrate models. Data to calibrate (i.e., train) the models were from validated assays conducted at NCTR laboratories. The compounds selected to train the models were selected based on providing uniform coverage of the diverse chemical structure space of chemicals that bind the receptors, as well as coverage of an activity range extending down to a million fold below that of the endogenous hormones. A large number of inactive chemicals were included in the training sets to enable models to be trained to distinguish active from inactive compounds.

Various EDKB predictive models have been developed using the many powerful commercial software packages that are now routinely applied in drug discovery and development. These comprise SAR, QSAR and chemometric methods. The development, application and validation of these models are discussed below in more detail.

We have developed a suite of models that when used together enable fast, high confidence predictions of ER-mediated estrogenicity for large databases, such as that for the U.S. Endocrine Disruptor Screening Program (EDSP).

Structure Activity Relationships for ER

Structure-activity relationships for estrogens has been well documented [Dodds, 1938 #501] to include the discovery of nonsteroidal estrogens and descriptions of the important structural features governing potency. Understanding structural requirements for a chemical to exhibit estrogenic binding affinity has been directly or indirectly applied to design drugs for human estrogen replacement therapy, and to develop computational models for rapidly screening and priority setting of testing estrogenic endocrine disruptors by regulatory bodies.

Fang [Fang, 2001 #1299] reported the structure-activity relationship (SAR) based on a total of 230 chemicals including both natural and xenoestrogens comprising the NCTR ER-binding training set. Their activities are generated using rat estrogen receptor competitive binding assay [Blair, 2000 #216], which cover a 1,000,000-fold range of binding. Focusing on identification of structural commonalties among diverse estrogens, the study shows how xenoestrogens structurally resemble endogenous estradiol and synthetic estrogen DES. Based on the SAR analysis, five distinguishing criteria were concluded essential for being an xenoestrogen using the endogenous estradiol as a template: 1) Phenolic ring: H-bonding ability mimicking 3-OH; (2) 3, 17 hydroxyl group O-O distance and H-bond donor mimicking 17b-OH; (3) precise space of steric hydrophobic centers mimicking 7a and11b steric substituents; (4) hydrophobicity; and (5) ring structure. The 3-position H-bonding ability is a primary requirement for estrogenic chemicals.

Hierarchical Models for Priority Setting Based on ER Binding

Legislation requiring the development and implementation of a strategy for screening and testing chemicals for estrogen, androgen and thyroid endpoints [EDSTAC, #545] led to the definition of a two-tiered, multiple-endpoint strategy by EPA's Endocrine Disruptor Screening and Testing Advisory Committee (EDSTAC) of which FDA is a participating member. This strategy incorporates more than 20 different in vitro and in vivo assays [Gray, 1998 #250]. As many as 87,000 chemicals may need to be screened for endocrine-disruption potential [Patlak, 1996 #526]. The large number of chemicals (87,000) and assays makes it difficult for each chemical to be run through these assay batteries in a reasonable time. There is a crucial need for priority setting to identify the chemicals most likely to possess endocrine disrupting activity for early entry into screening.

Priority setting using computational approaches is widely applied in the process of drug discovery. The objective of priority setting in pharmaceutical industry is to increase the chance of finding active compounds or "hits" that are more likely to be developed into "leads". Hence false positives are of great concern. In contrast, minimizing false negatives is critical for regulatory purposes because chemicals labeled as inactive are dropped into a lower priority category. For this purpose, we developed an integrated computational system [Shi, 2001 (in press) #226] that rationally combines different computational models into a sequential "Four-Phase" scheme according to the strength of each type of model . In Phase I [Hong, 2001 (in press) #227], several simple rejection filters or rules are used to exclude those chemicals that are most unlikely to exhibit estrogenic activity. Phase II [Hong, 2000 (in press) #227], uses three different types of models (structural alerts, pharmacophores, and classification methods) to make a qualitative activity prediction. In Phase III [Shi, 2001 #225], multiple quantitative structure-activity relationship (QSAR) models are used quantitatively to predict activity. In Phase IV, an expert system (not yet developed) is recommended to combine Phase II and Phase III predictions with exposure, fate and other data to set priorities. In this scheme, each Phase is used as a screen to reduce the number of compounds to be considered in the subsequent Phase. Therefore, these four Phases work in a hierarchical way to incrementally reduce the size of a dataset while simultaneously increase precision of predictions. Within each Phase, different complimentary models have been selected to represent key activity-determining structure features and to minimize the rate of false negatives.

The phases for the proposed integrated approach for priority setting is described below in more detail. Each phase contains a number of rules and/or models to estimate a compound's binding affinity:

Phase I: Filtering - A set of rejection filters are designed to significantly and with high confidence eliminate those chemicals with extremely low or no ER binding affinity. The key criterion in selecting these filters is to rule out as many true negatives as possible, while having an extremely low probability of passing a false negative.
Phase II: Active/Inactive assignment - Three different methods, i.e., structural alert, pharmacophore searching, and classification/clustering, are used in parallel in this phase to distinguish active from inactive compounds. Each method incorporates and weighs differently the various structural features that allow a chemical to bind the ER. In order to further reduce false negatives, a chemical predicted to be active by any of these methods is subsequently evaluated in Phase III. At Phase II, the combined output derived from the three approaches could be used for initial categorical priority setting.
Phase III: Quantitative predictions - A number of QSAR methods are used for quantitative prediction of the binding affinity of the compounds identified as active in Phase II [Tong, 1998 #175; Shi, 2001 #225]. Compounds with higher predicted binding affinity are given higher priority for further evaluation.
Phase IV: Rule-based decision-making system - In this final stage of our integrated priority setting approach, we propose to use a rule-based (or knowledge-based) decision-making system to foster definitive decision making. This system will be useful only after we incorporate our accumulated human knowledge and expertise, i.e., rules, into the system.

Hierarchical sequencing of the models allows faster models to be used to eliminate the majority of inactive chemicals with an extremely low rate of false negatives. The more time-consuming but more precise models can be used to refine predictions for an increasingly smaller number of remaining chemicals. The application of the more refined models further eliminates true negatives as well as false positives from earlier models.

Structure Activity Relationships for AR

The development of AR binding models is currently ongoing. The process parallels that carried out for the AR model development. The website will be updated to reflect this work once the results are published.

Model Development Process

Model Validation Program

A validation program is now underway via an Interagency Agreement between FDA/NCTR and EPA to assess 200 chemicals selected by EPA. These chemicals will be tested blind by the models, and the predictions then compared to assay results from an outside contract laboratory. The validation results will define whether and how model predictions are used in priority setting in the EPA's Endocrine Disruptor Screening and Testing Program.