• The selection of compounds with optimal pharmacokinetic properties, its stability or availability in biological systems.
QSAR methods are also categorized into following classes, based on the way by which the descriptor values are developed (figure 1):
0D-QSAR is achieved based on descriptors derived from molecular formula such as molecular weight, number of atoms, atom types, sum of atomic properties;
1D-QSAR (correlates activity with global molecular properties like pKa, solubility, log P, functional groups(substructures) etc);
2D-QSAR (correlates activity with structural patterns like connectivity indices, 2D-pharmacophores wiener index, etc., without taking into account the 3D-representation of these properties);
3D-QSAR (deals with the orientation of the molecules in space, correlates activity with non-covalent interaction fields “steric and electrostatic field” surrounding the molecules);
4D-QSAR (additionally including ensemble of ligand configurations in 3D-QSAR, by representing each molecule in different conformations, stereoisomer’s, orientations, tautomer’s, or protonation states.);
5D-QSAR (explicitly representing different induced-fit models in 4D-QSAR, as the manifestation and magnitude of the induced fit may vary for binding of individual molecule to target protein);
6D-QSAR (further incorporating different solvation models in 5D-QSAR, can be achieved explicitly by mapping parts of the surface area of the molecules with solvent properties); HiT QSAR (hierarchical quantitative structure–activity relationship technology) based on the Simplex representation of molecular structure (SiRMS) and its application for different QSAR tasks. The spirit of this technology is a sequential solution (with the use of the information obtained on the previous steps) to the QSAR problem by the series of enhanced models of molecular structure description [from one dimensional (1D) to four dimensional (4D)].
QSAR: Basic Steps Involved
The QSAR model development trail involves four steps:
(a) Data collection and descriptor calculation,
(b) Analysis of correlation between the descriptors and input data structures (2D/3D),
(c) Validation of models and
(d) Design and activity prediction of new molecules.
For the development of a QSAR model, series (con-generic series in some cases e.g. CoMFA/CoMSIA, 3D-QSAR methods) of molecules, defined by their structures, with known activity data are used as input data for calculation of molecular descriptors. For all the compounds, the activity data should be in same units of measurement (binding/functional/IC50/Ki). Ki value is preferred instead of the IC50 data, since it is independent of the substrate concentration. To develop a QSAR, a significant number of compounds are requisite to develop a meaningful relationship. It is widely accepted that 5 to 10 compounds are required for every descriptor in a QSAR. This does suggest that a 1 descriptor regression-based QSAR could be developed. However in an ideal world “many more” compounds are required to obtain statistically healthy QSARs.
The structure cannot be directly used for creating structure-activity mappings for reasons that they do not usually contain in an unambiguous form the information that relates to activity. To evade this barrier, rationally designed molecular descriptors convert the structure to well defined sets of numerical values that can be correlated directly with the activity. Depending on data set, amongst the thousands of different descriptors available, only a few are well correlated to the activity and hence, the use of excess number of descriptors may affect the interpretability of the final model. To avoid this problem in QSAR analysis, a wide range of techniques for automated thinning of the set of descriptors to the most informative ones are used. This is followed by the splitting of the dataset into training and test sets using different approaches of hierarchical and non-hierarchical clustering, which is based on many algorithms including sphere exclusion theory or simply activity ranking. However, according to some authors, such similarity based splitting methods may result in selection of a biased test set. Hence to overcome this, splitting may also be done based on random selection of subset molecules to generate test set data.
The training set compounds are then used for building a correlation between the descriptor values and the activity data. The resulting correlation is obtained using different chemometric tools (both linear and non-linear) like PLS, PCA, ANN, kNN etc and the model quality is judged using a variety of statistical parameters. Further, the model thus obtained undergoes both internal (leave-one-out method etc.) and external validation (test set prediction etc). The predictivity and reproducibility of the QSAR model is decided based on its ability to predict the test set molecules. A proper interpretation of a QSAR model exposes the key structural features obligatory for an improvement in the definite response parameter being modeled. Hence the developed model can be used further for design and activity prediction of new series of molecules.
QSAR is a widely used tool for developing relationships between the effects (activities and properties of interest) of a series of molecules with their structural properties. It is used in many areas of science. It is a dynamic area that integrates new technologies at a staggering rate.
Topics which I have discussed above were required to introduce you to QSAR. This is just the beginning; there is plenty to share about QSAR, it is not just a method where some steps are involved in developing model. I would like to discuss QSAR models and what they signify. I didn’t discuss it here since this is just an introduction. In future I will also discuss about ‘validation’. Validation of each step is important, if you don’t want to end up with an erroneous model. Even after model development, validation plays a vital role. Any QSAR study is as good as the model on which it is based.
In conclusion I would say that usefulness of QSAR as a technique relies on generation of more robust models, which is the next big challenge.
“Remember that all models are wrong; the practical question is how wrong do they have to be to not be useful.”
- George E. P. Box
Posts
