ConvPred: A deep learning-based framework for predictions of potential organic reactions

The traditional theory-based technologies on reaction prediction are not efficient due to their heavy dependence on human expertise and experience. To this end, this article proposes a framework for predictions of potential organic reactions based on reaction templates and two-dimensional convolutional neural network (2D CNN) model. The quantum mechanics-based sigma-profiles and the sub-molecular structure-based ECFP4 are used individually to encode chemical reactions. Using 605,753 patented reactions extracted from the USPTO 1976-2016 database and their generated counterparts, the 2D CNN models are trained to evaluate the likelihood of molecular transformations by learning the feature differences between reactants and products. The classification accuracies of the sigma-profiles-based model and the ECFP4-based model for the non-trained reactions are 97.881 and 99.593%. Challenging reactions from literature involving identification of chemo-, stereo-, and regio-selectivity are correctly predicted. Furthermore, a sigma-profiles-based visual reaction fingerprint is introduced to provide novel insights into the model interpretability.