Untitled

Abstract

A novel molecular descriptor (EVA) based upon calculated infrared range vibrational frequencies is evaluated for use in QSAR studies. The descriptor is invariant to both translation and rotation of the structures concerned. The method was applied to 11 QSAR datasets exhibiting both a range of biological endpoints and various degrees of structural diversity. This study demonstrates that robust QSAR models can be obtained using the EVA descriptor and examines the effect of EVA parameter changes on these models; recommendations are made as to the appropriate choice of parameters. The performance of EVA was found to be comparable in statistical terms to that of CoMFA, despite the fact that EVA does not require the generation of a structural alignment. Models derived using semiempirical (MOPAC AM1 and PM3) and AMBER mechanics calculated normal mode frequencies are compared, with the overall conclusion that the semiempirical methods perform equally well and both outperform the AMBER-based models.

Untitled

“…The EVA descriptor [1,2] is derived from fundamental molecular vibrational frequencies of which there are 3N-6 (or 3N-5 for a linear compound such as acetylene) for an N-atom structure. The frequency values are projected onto a linear bounded frequency scale covering the range 1 to 4,000 cm -1 and then smeared out, and therefore overlapped, through the application of Gaussian kernels to each and every frequency value.…”

“…In most cases the EVA models were found to be statistically entirely comparable to those obtained using CoMFA but without the difficulties associated with structural superposition. A detailed study with a benchmark steroid dataset [3] indicated that EVA can provide statistically robust QSAR models when this is judged by the scores from internal crossvalidation, random permutation tests and external test set prediction.…”

“…This is necessary because it means that each frequency (i.e., each part of the spectrum) is equally weighted prior to regression analysis. It has been found that the quality of the QSAR model very often is dependent upon the chosen σ and that the best σ to use can vary substantially [2,3]. The general approach [3] has been to generate many sets of EVA descriptors based upon a variety of σ and, on the basis of training set crossvalidation results, select a σ expansion term that is expected to provide an optimally predictive model for a previously unseen test set.…”

“…It has been found that the quality of the QSAR model very often is dependent upon the chosen σ and that the best σ to use can vary substantially [2,3]. The general approach [3] has been to generate many sets of EVA descriptors based upon a variety of σ and, on the basis of training set crossvalidation results, select a σ expansion term that is expected to provide an optimally predictive model for a previously unseen test set. The effectiveness of this model-selection method has been clearly demonstrated with a steroid dataset [3].…”

“…An incentive for the development of this new approach to EVA, referred to hereafter as EVA_GA, is that there are a number of datasets with which EVA has previously not performed well [2], either in absolute terms or relative to CoMFA -the reasons for this under-performance are not apparent. In addition to the potential for providing improved QSAR predictions, it may be the case that the use of localised σ improves the possibilities for interpretation of an EVA QSAR model (i.e., back-tracking to structure).…”