Deep Learning on Molecular Structures
Deep neural networks have been the focus of machine learning research for the past few years. They have mostly been applied to images, audio and text. Their usefulness for molecule data, however, is still relatively unexplored. Here it is still common to use predefined procedures to turn the molecular structure into a numerical vector. Representing the graph in a relatively raw form and letting the network learn how to generate useful features from this representation, has the potential to lead to a feature vector that performs better for the machine learning task at hand. In addition it can give insight into which parts of the structure had the most influence on the result.