FP2VEC

Abstract

The quantitative structure-activity relationship (QSAR) models are regression or classification models to predict the chemical properties of compounds. An exact prediction of QSAR models can save time and costs compared with actual experiments. For the prediction of QSAR model, the molecular featurizer, the numerical expression of a chemical compound is also important. Recently, the machine learning and deep learning techniques are widely used to develop new molecular featurizers to improve the prediction accuracy of QSAR model. Here we introduce the new method for the molecular featurizer, FP2VEC, inspired by the natural language processing techniques. Our new method can express the chemical compounds as a vector representation which is trained by a supervised learning method. And we built a QSAR model using a simple convolutional neural network to evaluate the prediction performance of the FP2VEC method. We evaluated the prediction performance of our model against four for the classification tasks and five datasets for the regression tasks. And we compared our model with other molecular featurizer models. On the classification tasks, our model showed the best prediction accuracy among the benchmark models on three out of four datasets. Also, our model implemented with multi-task learning method outperformed other the benchmark models. And on the regression tasks, our model showed the best performance two out of five datasets. Lastly, we tested the effect of the hyperparameters in our model, and some hyperparameters influenced to the prediction accuracy significantly. As a result, our new molecular featurizer based on NLP techniques provides more useful information and improved the prediction accuracy of QSAR prediction compared with the previous methods.