Cell-specific imputation of drug connectivity mapping with incomplete data

Motivation: Drug repositioning 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 compounds whose impact on expression in a collection of cells reverses the disease9s impact on expression in disease-relevant tissues. The high throughput LINCS project has expanded the universe of compounds 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 repurposing drugs this way despite missing data, we compared collaborative filtering with either neighborhood-based or SVD imputation methods to two naive approaches via cross-validation. Results: Methods were evaluated for their ability to predict drug connectivity despite missing data. Predictions improved when cell type was taken into account. Neighborhood-based collaborative filtering was the most successful method, with the best improvements in non-immortalized primary cells. We also explored which classes of compounds are most and least reliant on cell type for accurate imputation, and we identified connections between related compounds even when many were not measured in the relevant cells. We conclude that even for cells in which drug responses have not been fully characterized, it is possible to identify unassayed drugs that reverse in those cells the expression signatures observed in disease.