Learning functional group chemistry from molecular images leads to accurate prediction of activity cliffs

Advances in image analysis through deep learning have catalyzed the recent use of molecular images in chemoinformatics and drug design for predictive modeling of compound properties and other applications. For image analysis and representation learning from molecular graphs, convolutional neural networks (CNNs) represent a preferred computational architecture. In this work, we have investigated the questions whether functional groups (FGs) and their distinguishing chemical features can be learned from compound images using CNNs of different complexity and whether such knowledge might be transferable to other prediction tasks. We have shown that frequently occurring FGs were comprehensively learned, leading to highly accurate multi-label FG predictions. Furthermore, we have determined that the FG knowledge acquired by CNNs was sufficient for accurate prediction of compound activity cliffs (ACs) via transfer learning. Re-training of FG prediction models on AC data optimized convolutional layer weights and further improved prediction accuracy. Through feature weight analysis and visualization, a rationale was provided for the ability of CNNs to learn FG chemistry and transfer this knowledge for effective AC prediction.