Influence of spatial information coding into 2D and 3D descriptors for QSAR modelling purposes

Abstract

The multivariate image analysis applied to QSAR (MIA-QSAR) is a technique based on the treatment of bidimensional images resulting from the projections of perfectly congruent, non-optimized geometries. It stands out for being a methodology that balances simplicity and efficiency in the generation of prediction models of biological/physicochemical properties. Because MIA-QSAR is a 2D technique, it does not efficiently encode spatial information in its molecular descriptors. In this sense, and keeping in mind the key role of 3D information for modeling and describing biological/physicochemical properties of molecules, the present work aims to investigate different strategies to encode and model this type of information in MIA-QSAR descriptors, as well as to evaluate the role of conformation in an originally tridimensional QSAR approach. Three different sources of descriptors have been proposed to codify 3D information into the MIA-QSAR descriptors: (I) images of 2D projections of compounds with previously optimized geometries; (II) images of “molecular slices” obtained after scanning molecules, with optimized geometries, along one of the cartesian axes; and, (III) images from the front, right and top faces of chemical structures, with optimized geometries, placed inside a theoretical box. For data modeling, two robust multivariate regression tools were used: for 2D projection descriptors, the support vector machine applied to regression (SVR) method was employed; for the other two strategies, the multilinear partial least squares (N-PLS) method was chosen. The three routines were applied to three different groups of compounds, a series of molecules with activity against the hepatitis C virus (anti-HCV), another with action against the coronavirus that causes severe acute respiratory syndrome (SARS-CoV), and a group with anti-HIV activity (human immunodeficiency virus). As a result, high quality parameters for both internal and external validation were achieved in all three strategies, and the statistical results of correlation were at least similar to those earlier reported for these series of compounds. Nevertheless, the risk of chance correlation could not be excluded as demonstrated by y-randomization tests. Accordingly, traditional MIA-QSAR method that uses perfectly congruent, non-optimized geometries of pharmacophoric substructures as images is still more efficient than the attempts to incorporate 3D information in the modelling. To evaluate the role of conformational information in an originally 3D-QSAR technique, one compared models built with variables codifying tridimensional aspects fully described, obtained from chemical structures previously docked in their biological target, with descriptors in which this type of information is either suppressed (flat structures) or only partially described (chemical structures with computationally optimized geometries). As a result, the validation parameters indicated that the robustness of the QSAR models seems to be more related to the alignment aspects of the structures than to the level of detail of tridimensional aspects encoded by the molecular descriptors.