Prediction of Organic Chemical Reactions Using Cyclical Learning Rate Based Long-Short Term Memory

The organic chemical reaction is a transformation of one or more organic components into an innovative molecule through the formation and breaking of chemical bonds. Through the data-driven techniques leads to sub-optimum results through making prediction on data distributions. In this paper, the Cyclical Learning Rate (CLR) based Long-Short Term Memory (LSTM) is proposed for the prediction of organic chemical reactions. The USPTO-50K dataset is utilized which comprises 50016 reactions through the accurate atom-mapping for every reactions. In data preprocessing, the actual datasets are comprised of reaction SMILES, so each reaction SMILES sequence for extract each reactant and product to produce corresponding template and class labels. At last, the integrated features are fed into Fully Connected (FC) and SoftMax layer to obtain the prediction output. By cyclically increasing and reducing the learning rate, it enables the LSTM model to discover various learning rate during training time. The CLR based LSTM attained better accuracy of 63.4% for top 1, 81.3% for top 3,88.7% for top 5 and 93.8% for top 10 when compared to existing methods like Local Reactive and Global Attention and Convolutional Neural Network.