Transfer Learning as Tool to Enhance Predictions of Molecular Properties Based on 2D Projections

Images of molecules are widely used to predict molecule properties in teaching and chemical research. A trained chemist can easily derive molecule properties by analyzing its structure and evaluate its functional groups. To predict, for example, the water solubility of an organic compound a chemist would intuitively count the number of polar groups, consider the size of the molecule, and estimate the water/molecule interaction by counting the number of H-bond donors and acceptors. Therefore, 2D molecule representations and their directly accessible features should provide enough information to predict the molecule's structure-dependent properties. To support this thesis, different image-based machine learning approaches as dense neural networks, convolutional neural networks, clustering, data augmentation, and transfer learning are compared and evaluated in this work. The influence of the image size as well as the network size is discussed. Finally, a simple yet effective dense neural network trained on expert preselected, visually accessible features, is presented and its efficiency and comparability to other more complex methods are demonstrated.