Improving cell-specific drug connectivity mapping with collaborative filtering

MotivationDrug re-positioning allows expedited discovery of new applications for existing compounds, but re-screening vast compound libraries is often prohibitively expensive. "Connectivity mapping" is a process that links drugs to diseases by identifying drugs whose impact on expression in a collection of cells most closely reverses the diseases impact on expression in disease-relevant tissues. The high throughput LINCS project has expanded the universe of compounds, cellular perturbations, and cell types for which data are available, but even with this effort, many potentially clinically useful combinations are missing. To evaluate the possibility of finding disease-relevant drug connectivity despite missing data, we compared methods using cross-validation on a complete subset of the LINCS data. ResultsModified recommender systems with either neighborhood-based or SVD imputation methods were compared to autoencoders and two naive methods. All were evaluated for accuracy in prediction of both expression signatures and connectivity query responses. We demonstrate that cellular context is important, and that it is possible to predict cell-specific drug responses with improved accuracy over naive approaches. Neighborhood-based collaborative filtering was the most successful, improving prediction accuracy in all tested cells. We conclude that even for cells in which drug responses have not been fully characterized, it is possible to identify drugs that reverse the expression signatures observed in disease. Contactdonna.slonim@tufts.edu Supplementary informationbcb.cs.tufts.edu/cmap