From Quantification of Activity Landscape Topographies to Activity Cliff Analysis

Abstract

Machine learning or multivariate statistical modeling is frequently applied in the chemical informatics domain to analyze the relationships between molecular structures and their target activities/properties. Structure-activity relationship (SAR) properties and key determinants have been vastly investigated with the help of simple linear to complex 2D graphs, 2D molecular fingerprints, and 3D conformation representations. SAR characteristics can be graphically visualized and analyzed with the help of 3D activity landscape (AL) models. 3D AL models help in a straightforward interpretation of SARs by characterizing them into smooth valleys and rugged mountainous regions, reminiscing SARs as a geographical map. 3D ALs graphical nature makes them readily convertible to image data, thus amenable for image analysis. 3D AL models graphical representation of important SAR characteristics, and their readily convertibility into image formats permit SAR analysis based on image processing approaches. However, 3D ALs representation has thus far been qualitatively analyzed and limited to very few quantitative and predictive analyses. Convolutional neural networks (CNNs) are currently gaining increasing attention in chemical informatics domain to learn key structural features and derive predictions based on compound image representations. CNNs have the potential to automatically extract distinct spatial and temporal features from images, audio, and video data, which is preferable as compared to classical machine learning modeling. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) gave birth to many complex and deep CNN model architectures that have solved several other complex image analysis problems afterward. This dissertation presents contributions that advance the application of CNN modeling for SAR and activity cliff analysis and the application of recent image processing approaches to develop methodologies to quantify SAR characteristics numerically.