Computer Based Strengths and Weaknesses

Of all of TiPED’s five testing tiers, the computer-based modeling is the least developed.  In an ideal world, one would be able to enter a molecular structure into a computer and have an instant read on its potential endocrine activity.  Unfortunately, such tools are not yet readily available and what tools exist are not accurate enough to stand alone.  That said, progress is being made toward this ultimate goal. For instance:

Making predictions about how a molecule will act based on quantitative structure activity relationships or QSAR:

Although it is potentially a useful statistical tool, obtaining a meaningful Q/SAR predictive model on toxicity is problematic, and depends on several factors, including the quality and availability of biological data, the statistical methods employed, and the choice of descriptors. A useful Q/SAR model would incorporate the following characteristics:

1) Include a training set comprised of a sufficient number of molecules that cover the range of properties to be predicted by the model.

2) The number of compounds in the training set should be far more numerous (at least 5 to 10 fold) than the number of non-correlated descriptors used to calculate the model. Furthermore, the descriptors should be biophysically relevant to the property being predicted.

3) The model should be applicable to novel compounds and allow for mechanistic information related to the endpoint of interest.

4) Preferably, the simplest model should be selected.

For the purposes here, a chemist should consider the following limitations of the Q/SAR approach when selecting a Tier 1 method to predict EDC potential:

  • The “SAR Paradox”, the fact that molecules of similar structure often have very dissimilar biological activity[16].
  • Each Q/SAR model predicts a specific endpoint, and only for chemicals with the identical mechanism.
  • Q/SAR models do not perform well with chemical structures outside the training set.
  • Most nuclear receptors have not been the focus of Q/SAR modeling, and there almost certainly are receptors yet to discover. Existing Q/SAR models predict only a subset of potential endocrine-activity and as such are insufficient.
  • Q/SAR models do not predict whether the compound agonizes or antagonizes a receptor.
  • Care must be taken to avoid deriving an over-fitted model (e.g. one that describes random error or noise, rather than an underlying relationship) and generating useless interpretations of structural/molecular data.

In sum, while Q/SAR models currently can be used as statistical tools for broad statements of probability they are not yet sufficiently developed for predictive toxicology, especially for endocrine disruption; additional tools must be used to provide a fuller picture.

Modeling of Biological Activity (Pocket Modeling, Molecular Docking):

The simplest way to think about a molecule and its receptor is to picture them as a lock and key, with a caveat that both of them are somewhat flexible.  In a molecular docking model, the goal is to determine the correct orientation and adjustments of these two components. Specifically, molecular docking predicts the preferred orientation a molecule will adopt when bound to another molecule (i.e. the receptor) to form a stable complex. This information can be used to predict the binding affinity, or strength of association between the two molecules. Because the relative orientation of two molecules influences whether agonism or antagonism of the receptor results from their interaction, this method is useful for determining what type of signal a novel chemical is predicted to generate at the receptor. The limitation of this approach is that the molecular docking method requires an available crystal structure of the ligand-binding domain of interest, or at least of its close relative, as well as understanding of the domain’s flexibility, and structures being altered by residence in different cellular locations, such as plasma membrane vs. aqueous compartments.

The main approach used by scientists that study molecular docking simulates the actual docking process, whereby the ligand moves into position within the receptor’s active site following a series of rigid body transformations and internal changes to the ligand structure, such as torsion angle rotations, as well as changes in the binding pocket structure [12]. Unlike simple comparisons of the complementarity of receptor and ligand shapes, simulation approaches can incorporate both ligand and receptor flexibility into the model, thus it is more reflective of what actually happens during ligand-receptor interactions. A disadvantage of this approach is that it is more time-consuming.

Molecular docking modeling tools have been developed in connection with pharmaceutical chemistry and are now being adapted to predict endocrine disruption potential.  Initial studies have demonstrated the acute accuracy of the tool, e.g. accurately modeling the interaction of polybrominated diphenyl ethers (PBDEs) with the ER[17, 18] and AR[19], as well as preliminary studies of a panel of NRs with crystallographic structures[20].  Recent tests of PPARγ models demonstrate the very strong (at close to 100% accuracy) discriminating ability of the docking models.  As this particular tool is further developed and refined, its utility in predicting EDCs will become extremely valuable as part of this tier in the TiPED toolbox.