Untitled

Abstract

The EVA molecular descriptor derived from calculated molecular vibrational frequencies is validated for use in QSAR studies. EVA provides a conformationally sensitive but, unlike 3D-QSAR methods such as CoMFA, superposition-free descriptor that has been shown to perform well with a wide range of datasets and biological endpoints. A detailed study is made using a benchmark steroid dataset with a training/test set division of structures. Intensive statistical validation tests are undertaken including various forms of crossvalidation and repeated random permutation testing. Latent variable score plots show that the distribution of structures in reduced dimensional space can be rationalized in terms of activity classes and that EVA is sensitive to structural inconsistencies. Together, the findings indicate that EVA is a statistically robust means of detecting structure-activity correlations with performance entirely comparable to that of analogous CoMFAs. The EVA descriptor is shown to be conformationally sensitive and as such can be considered to be a 3D descriptor but with the advantage over CoMFA that structural superposition is not required. EVA has the property that in certain situations the conformational sensitivity can be altered through the appropriate choice of the EVA sigma parameter.

The EVA spectral descriptor

“…are not considered) but rather it is to apply a smearing function such that vibrations at slightly different frequencies in different compounds can be compared with one another. As such the results obtained with EVA QSAR are usually dependent upon the chosen kernel width (σ) [20][21][22] since this parameter determines whether or not, and the extent to which, proximal kernels overlap. A general approach for choosing an appropriate Gaussian σ is described below together with a detailed explanation of how the sampling resolution (determined by L) should be selected.…”

“…Given that the descriptor variance depends upon these factors it follows that any variance-based method such as PLS is sensitive to the chosen Gaussian σ. It is indeed the case that optimal σ (as judged by the resulting PLS scores) can be identified for particular data sets [20,21] although the sensitivity to σ is a data set-dependent feature. The much discussed "benchmark" steroid data set [26], for example, is particularly sensitive to σ (as demonstrated in Figure 1) [21].…”

“…It is indeed the case that optimal σ (as judged by the resulting PLS scores) can be identified for particular data sets [20,21] although the sensitivity to σ is a data set-dependent feature. The much discussed "benchmark" steroid data set [26], for example, is particularly sensitive to σ (as demonstrated in Figure 1) [21]. In this example, models were obtained for a range of σ from 1 to 25 cm -1 and LOO crossvalidation, fitted modelling and prediction performed for each descriptor set; PLS model dimensionality was chosen on the basis of the first SE cv -minimum.…”

“…There has thus been much interest in either tackling the alignment issue head-on [9,[11][12][13][14][15] or in seeking alternative molecular descriptors that are both sensitive to 3D-structure but that do not require structural superposition [11,13,14]. EVA [18][19][20][21][22][23] is one example of such a descriptor, based as it is upon molecular vibrations the characteristics of which are, in the absence of an external modifying influence such as a receptor, invariant to rotation and translation of the structures concerned. However, whilst EVA removes the need for superposition the method is sensitive to 3D structure although not to such an extent as a "true" 3D method such as CoMFA.…”

“…Nonetheless, it has been demonstrated that it is beneficial to "match" conformations across a dataset where possible rather than using randomly or arbitrarily selected 3D structures [21].…”