Optibrium

No. The P450 models differ from the QSAR models, being based on simulation of the chemical reaction mechanisms which lead to the formation of metabolites. Although experimental data are used to tune the parameters of the model and validate the results, the form of the underlying model is not based on an empirical fit to a training data set and this gives greater transferability across a wide range of chemistry without loss of accuracy. Accurate modeling of the chemical reactions requires quantum mechanical simulations, which are much more computationally expensive than the descriptor calculations employed by QSAR models. Consequently, the P450 metabolism models are significantly slower, taking several minutes per compound. However, identifying the most likely cause of metabolic instability for a compound can help to guide chemical modifications aimed at reducing the vulnerability.

We use an adapted version of MOPAC97 employing the AM1 method. As a semi-empirical method, AM1 is known to exhibit some systematic errors. Therefore, post hoc corrections are applied which have been derived from extensive ab initio calculations.

Initial 3D geometries are generated using Corina (developed by Molecular Networks) and these are then optimised further within MOPAC97. Unconstrained geometries are used, as experimental analysis has shown that the results are relatively insensitive to the geometry provided a good local minimum is used.

Yes, the CSL values are reported with an uncertainty in prediction and can be used in the scoring profile.

In addition to the vulnerability of individual sites on a molecule, whether a compound is metabolised, and the overall rate of metabolism, will depend upon many factors in the P450 catalytic cycle in addition to site lability, including the affinity of the molecule for the enzyme, the rates of reduction and rates of decoupling via peroxide formation. These are currently unknowns and hence metabolic rate, as opposed to lability, cannot be predicted in general. However, for individual chemical series typically only a small number of these factors dictate the rate of product formation. In this case it is possible to build local models, using the CSL as a descriptor.

No, these models do not predict the most likely isoform responsible for metabolism.

The accuracies of the models for each isoform when tested on independent test sets are shown below:

Results of the P450 Regioselectivity Models on independent test sets

* N = number of compounds in independent test set.
** Observed metabolites in the top 3 predicted sites or >10% predicted metabolism.

Only 14 of the 22 molecules that form the top 20 drugs in this study are metabolised by Cytochrome P450 enzymes. These 14 drugs produce a total of 22 major metabolites of which 19 are correctly identified (86%), and a total of 6 minor metabolites of which 5 are correctly identified (83%). The results of the study are summarised below.

Summary of Results for Top20 Drugs* Using Regioselectivity Models

*By sales for 2004. Source: Med. Ad. News May 2005.