Deepcellstate: An Autoencoder-Based Framework For Predicting Cell Type Specific Transcriptional States Induced By Drug Treatment

Drug treatment induces cell type specific transcriptional programs, and as the number of combinations of drugs and cell types grows, the cost for exhaustive screens measuring the transcriptional drug response becomes intractable. We developed DeepCellState, a deep learning autoencoder-based framework, for predicting the induced transcriptional state in a cell type after drug treatment, based on the drug response in another cell type. Training the method on a large collection of transcriptional drug perturbation profiles, prediction accuracy improves significantly over baseline and alternative deep learning approaches when applying the method to two cell types, with improved accuracy when generalizing the framework to additional cell types. Treatments with drugs or whole drug families not seen during training are predicted with similar accuracy, and the same framework can be used for predicting the results from other interventions, such as gene knock-downs. Finally, analysis of the trained model shows that the internal representation is able to learn regulatory relationships between genes in a fully data-driven manner.

Author summaryA large number of gene expression profiles across different cell types are available, however many drug-cell combinations have not been profiled. Motivated by the need for accurate methods for prediction of cell type specific drug responses, we developed DeepCellState, a deep learning framework, with the goal of predicting the response in a given cell type based on the response in another cell type. Training the method on the largest available database for transcriptional response to drug perturbations, LINCS, we observed that the method can predict with high accuracy the cell type specific response for treatment with drugs not seen by the method, with improved accuracy as we generalized the method from two to multiple cell types. Encouragingly, the method performed well even when the response from completely unseen cell types were used as input. We further confirmed the robustness of our results through validation with data independently generated on other expression profiling platforms. Analysis of the learned models revealed although the training is completely data driven and uses no prior knowledge about regulatory relationships between genes, the network itself is biologically interpretable and captures interactions between transcription factors and the targets they regulate.