Different approaches to encode and model 3D information in a MIA-QSAR perspective

Abstract

Tridimensional information is a fundamental aspect for modelling and explaining biological/physicochemical properties. In this sense, the goal of this study was to explore different approaches for encoding this type of information into MIA-QSAR (Multivariate Image Analysis applied to Quantitative Structure-Activity Relationships) descriptors and to effectively model these new features. Originally, MIA-QSAR is a technique based on the treatment of 2D images of molecules. The approaches explored in this work were: (I) the use of 2D image projections of computationally optimized molecular geometries as a source of 3D information for a powerful machine learning method (support vector machine applied to regression); (II) the use of slice images obtained from the optimized molecules placed inside a theoretical box as a source of 3D descriptors for a multi-way regression method (trilinear PLS); and (III) the use of images viewed from different faces of the above box as an alternative source of 3D MIA-QSAR descriptors. These strategies were applied in three different data sets comprising anti-HCV, anti-SARS-CoV, and anti-HIV compounds. Satisfactory parameters for both internal and external validation were achieved in all three models, 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. Whereas the traditional MIA-QSAR method, that uses perfectly congruent, non-optimized geometries of pharmacophoric substructures as images, is more efficient than 3D MIA-QSAR, the latter uses tridimensional digital objects as descriptors for the first time in QSAR for regression purposes.