Abstract PR11: Reference pharmacologic class analysis for Connectivity Map discovery

Abstract

Abstract The Connectivity Map (CMap) is a database of 1.4 million transcriptional profiles from cell lines treated with small molecule or genetic perturbations. This resource promises to serve as a comprehensive “lookup table” for biology by linking drugs to genes and pathways. However, distinguishing true similarities from random noise within such a large, multidimensional dataset presents significant challenges to result interpretation and hit prioritization. To address these challenges, we took advantage of the fact that different CMap drugs may share the same molecular target or mechanism of action. We hypothesized that grouping drugs into reference pharmacologic classes would yield dramatically more crisp and readily interpretable query results. In order to successfully group drugs by mechanism of action, standardized drug annotations were required. More than 2,000 CMap compounds were mapped to their known targets by comparing chemical structures to entries in existing public databases or by manual searches using the scientific literature. Target information was stored as HUGO gene symbols. Next, targets were manually grouped into known pathways. The compound signatures of each class were assessed for internal consistency using a gene set enrichment-based approach known as introspection, and outlier compounds were eliminated. Using this approach, we identified and validated more than 50 reference pharmacologic class lists (PCLs), that include agonists or antagonists of a broad range of target types, such as GPCRs, nuclear hormone receptors, kinases, heat shock proteins, ribosomal proteins, biosynthetic enzymes, and others. Reference PCLs were evaluated through several approaches. First, a number of PCLs have strong connections to corresponding genetic treatments. For example, the MEK inhibitor PCL strongly matches MEK knockdown and is anti-correlated with RAF1 overexpression. Second, novel compounds with structures similar to known drugs connect to expected groups: for example, a derivative of betamethasone connects to corticosteroids, and a derivative of fasudil connects to ROCK inhibitors. Finally, we have observed multiple examples of unannotated compounds for which we can retrospectively find supporting evidence for the predicted mechanism. Applications of PCL analysis include screening for bioactivity, predicting novel compound mechanism of action, and assessing pathway modulation status from virtually any input query. We have developed a web application to enable scientists to easily view and explore the PCL connectivity results. Planned future experiments include testing reference compounds at multiple doses and adding additional pharmacologic classes of disease relevance. Citation Format: Steven M. Corsello, Rajiv Narayan, Joshua Gould, Ted E. Natoli, Xiaodong Lu, Aravind Subramanian, Todd R. Golub. Reference pharmacologic class analysis for Connectivity Map discovery. [abstract]. In: Proceedings of the AACR Special Conference on Computational and Systems Biology of Cancer; Feb 8-11 2015; San Francisco, CA. Philadelphia (PA): AACR; Cancer Res 2015;75(22 Suppl 2):Abstract nr PR11.