QSAR Modeling Based on Conformation Ensembles Using a Multi-Instance Learning Approach

Abstract

Modern QSAR approaches have wide practical applications in drug discovery for designing potentially bioactive molecules. If such models are based on the use of 2D descriptors, important information contained in the spatial structures of molecules is lost. The major problem in constructing models using 3D descriptors is the choice of a putative bioactive conformation, which affects the predictive performance. The multi-instance (MI) learning approach considering multiple conformations in model training could be a reasonable solution to the above problem. In this study, we implemented several multi-instance algorithms, both conventional and based on deep learning, and investigated their performance. We compared the performance of MI-QSAR models with those based on the classical single-instance QSAR (SI-QSAR) approach in which each molecule is encoded by either 2D descriptors computed for the corresponding molecular graph or 3D descriptors issued for a single lowest energy conformation. The calculations were carried out on 175 data sets extracted from the ChEMBL23 database. It is demonstrated that (i) MI-QSAR outperforms SI-QSAR in numerous cases and (ii) MI algorithms can automatically identify plausible bioactive conformations.

3D chemical structures allow robust deep learning models for retention time prediction

“…46 Accordingly, each step during training, we randomly oriented voxel grids, and for the testing, we averaged the predicted retention times over 24 possible orientations of the input voxel grid. 47 We observed $10(1.6%) seconds improvement in the RT prediction due to random orientation during training and $1.5(0.3%) seconds additional improvement due to the averaging over the 24 predictions corresponding to different grid orientations. To evaluate the robustness with respect to the different conformations of a molecule, we calculated the normalized standard deviation from the predictions obtained for four different conformers of each molecule m in the train and test sets:…”

Physics-Inspired Structural Representations for Molecules and Materials

“…Nevertheless, there is growing evidence that the use of descriptors incorporating information on the 3D geometries can improve the accuracy of QSPR models, especially for difficult cases that involve very flexible molecules, , as well as for data analytics approaches for materials informatics . One of the core challenges in these efforts is the determination of the conformer geometries that should be used to evaluate the 3D descriptors, an operation for which several strategies have been explored to enhance the accuracy of QSPR models. , As we briefly discussed in section , one of the most promising research directions involves combining the high fidelity of density-based representations with a well-principled construction of ensembles of features. This is still a very active subject of research, with very encouraging results having been recently demonstrated for the prediction of the solubility of small molecules, the computational screening for antiviral drugs, and the prediction of enantioselectivity of organocatalysts …”

Is conformation relevant for QSAR purposes? 2D Chemical representation in a 3D‐QSAR perspective

“…However, it has been debated whether a single conformation or even the method for selection of relevant conformations would provide reliable three-dimensional information 5 ; therefore, a multiinstance learning approach has been proposed. 6,7 In turn, our question goes beyond this: which relevant conformational information do the 3D methods provide that 2D approaches should inform? Multivariate image analysis applied to quantitative structureactivity relationships (MIA-QSAR) is an essentially 2D-QSAR technique, in which the MIA molecular descriptors correspond to pixels of chemical structure images plotted in a blackboard.…”

“…According to these methods, the molecules of a data set are placed inside a three‐dimensional lattice and the fields (e.g., steric and electrostatic) are sampled at the intersections of this grid box. However, it has been debated whether a single conformation or even the method for selection of relevant conformations would provide reliable three‐dimensional information 5 ; therefore, a multi‐instance learning approach has been proposed 6,7 . In turn, our question goes beyond this: which relevant conformational information do the 3D methods provide that 2D approaches should inform?…”