Abstract Since most compounds do not induce changes in the transcriptomic levels of their target proteins in vivo, traditional gene set enrichment analysis methods can only retrieve downstream differentially expressed genes, which offer little hints to their targets. To address this problem, we proposed a graph convolutional network-based drug “on-target” pathway prediction algorithm, GDOP, which can predict small pathways that contain target gene through the power of deep learning algorithms. Our model receives as input structural information and biological characteristics (gene expression profiles) of molecules. After being trained on the publicly available LINCS data set, GDOP showed better generalization ability, reaching an AUC-ROC of 0.89 and an averaged Top10 accuracy of 0.63 on the test set. Besides, demonstrated that GDOP was able to use RNA-Seq data as input and achieved accuracy prediction results.

GDOP: A graph convolutional network-based drug “on-target” pathway prediction algorithm