CDDTR: Cross-Domain Autoencoders for Predicting Cell Type Specific Drug-Induced Transcriptional Responses

Gene expression profiles play an important role in biology and translational medicine. Due to the vast combinations of drugs and cell lines, the exhaustive measurements of transcriptional perturbation responses become challenging. There are some computational methods to predict drug responses, but the mapping between the drug responses of different cell lines is largely overlooked. We propose CDDTR, a cross-domain autoencoders based method, that uses the drug induced response of one cell line to predict the response of another cell line. For the perturbation profiles of 1570 drugs on seven cell lines in Phase II LINCS-L1000 data, the Pearson correlation coefficient (PCC) value of CDDTR achieves significant performance improvements compared to other state-of-the-art methods, including DeepCellState, VAE and MAGAN. The latent features of transcriptional perturbations of A375 and PC3 cells extracted by 10-fold cross-validation have a 0.663 PCC, revealing the competence of CDDTR to predict the cross-cell type responses. By integrating perturbations from multiple cell lines and incorporating pre-training, the predictive performance of CDDTR can be further improved. Source code is available at https://github.com/hanjunbaba/CDDTR .