Development of a deep learning model for predicting the effects of small molecules on gene expression levels

Abstract

Transcriptome has served as an important systems biology tool for examining the physiological status of a cell, and has been applied to a wide range of fundamental and application studies. As an increasingly greater volume of transcriptome profiles are being generated, they can serve as a useful resource for building a predictive model. Developing such a predictive model can be useful when examining transcriptome profiles for a large number of cells under varied culture conditions, which can be cost and time prohibitive. To this end, we develop a deep learning-based model, namely DeepCGR, which predicts the effects of a given compound on the expression level of a given gene in a human cell line. DeepCGR takes a protein sequence and structural information of a small molecule as inputs, and classifies whether the input protein is upregulated or downregulated in response to a given compound. In addition, z-score of the expression level of the input protein is also generated as an output, which allows the regression analysis. This model was developed for 15 different human cell lines, which showed the average AUC and F1 score of 0.879 and 0.425, respectively, for the upregulation, and 0.860 and 0.388 for the downregulation. The mean squared error was 0.837 in the range of -10 to +10, and the Spearman’s correlation coefficient was 0.739 in the z-score> ±1.5. DeepCGR will lay the ground for high-throughput transcriptome analysis of a large number of the combinations of cell lines and small molecules, and will be particularly useful for drug screening where the effects of a large number of compounds need to be examined.