Pushing the boundaries of 3D-QSAR

Abstract

Based primarily on further studies of a collection of eleven publications reporting fifteen successful 3D-QSAR relations, several phenomena are preliminarily described. The RMS error of 133 ligand binding energy predictions based on these successful 3D-QSARs is 0.75 kcal/mole, which compares favorably to the prediction accuracies of approaches that include the receptor. A similar result is obtained when topomer alignments are substituted for those published, with seemingly profound implications for the future of 3D-QSAR. The "alignment-averaged" molecular properties, log P and molar refractivity, have very little correlative power for these data sets, either alone or in combination with the 3D-QSAR field descriptors. The q (2 )metric for the number of PLS components necessarily tends to discard any unique or unconfirmed SAR information. Large drops in q (2) are thus to be expected whenever such unique information is first encountered. Predictive r (2) values from an exploratory new "series trajectory" analysis of these 3D-QSAR though highly variable do not differ much from their q (2) values, a phenomenon that seems to encourage prediction even when there are so few structures underlying a 3D-QSAR so that almost all information is unique.

Three-dimensional QSAR analysis and design of new 1,2,4-oxadiazole antibacterials

“…The only way of validating predictive power reliably is by the use of an external test set. 27, 29, 30 The activity values for the predictive compounds (the test set) gave an r 2 pred of 0.55, thus indicating a reasonable predictive power of the model. This predicted r 2 assesses the robustness of the QSAR model.…”

Estimation of influential points in any data set from coefficient of determination and its leave-one-out cross-validated counterpart

“…PRESS is a sum of squares of residual quantities. If 17 we accept the reasonable assumption of OLS regression that the PRESS residuals are not 18 serially correlated and they are normally distributed, their sum of squares shows χ DFFIT-2. We did not detect reliable correlation for the method called HD, and we got 7 negative correlation for the COVRATIO one.…”

Computer-aided drug design: the next 20 years

“…One notable variation of that technique emerged about 10 years ago, CoMSiA (Comparative Molecular Similarity Analysis) [22], but with that exception there has been surprisingly little innovation in this area of empirical 3D model construction. The vicissitudes of CoMFA are well-acknowledged [23], and the need for empirical 3D models in the area of ADMET are strongcases like hERG where one can readily construct homology models are the exception, not the rule. Just as the invention of CoMFA was catalyzed by the invention of partial-leastsquares as a data analysis tool, advances in empirical 3D model construction may be tied to Outlook 2.2, an improvement in techniques for learning from data.…”