The Broad Institute: Screening for Dependencies in Cancer Cell Lines Using Small Molecules
Using cancer cell-line profiling, we established an ongoing resource to identify, as comprehensively as possible, the drug-targetable dependencies that specific genomic alterations impart on human cancers. We measured the sensitivity of hundreds of genetically characterized cancer cell lines to hundreds of small-molecule probes and drugs that have highly selective interactions with their targets, and that collectively modulate many distinct nodes in cancer cell circuitry. Using robust analytical methods, we connected sensitivity measurements to genetic and cellular features of the cell lines in order to identify dependencies conferred by specific genotypes. We developed a growing resource, the Cancer Therapeutics Response Portal (CTRP), to provide analysis results and visualizations of statistically significant connections and to serve as a hypothesis-generating resource for the cancer biology community. Additionally, we made available all primary data such that it can be re-analyzed to yield further hypotheses as additional computational approaches and deeper genetic and epigenetic characterization of the cancer cell lines become available. Our hope is that insights developed from CTRP, first based on cell line models of cancer and then substantiated in more complex environments, will yield clinical predictions of how patients will respond to novel types of targeted therapies and to accelerate the discovery of new genetically matched medicines.
Cancer Therapeutics Response Portal (CTRP v1, 2013) dataset: 355 small molecules; 242 cancer cell lines (CCLs). See Basu, Bodycombe, Cheah et al. for complete experimental details.
Cancer Therapeutics Response Portal (CTRP v2, 2015) dataset: 545 small molecules and select combinations; 907 cancer cell lines (CCLs). See Seashore-Ludlow et al. for experimental details.
CTRP reference implementation core code: https://github.com/remontoire-pac/ctrp-reference
Informer Set Assembly
Small molecules were selected to perturb potential targets and processes on which cancer cells may become dependent, including but not limited to oncogenes/tumor suppressors, DNA-damage response, reactive-oxygen species metabolism, electrophile-stress response, protein degradation, hypoxic-stress response, mitotic-stress response, survival/apoptosis, nutrient metabolism, nutrient-stress response, chromatin modification, and other major signaling pathways (e.g., PI3K/mTOR; NFkB; others). We performed an evaluation of the probe-development literature including: seminars, peer-reviewed journals, NIH Molecular Libraries Initiative Probe Reports, and patents. Many of our compounds were accessed through synthetic organic chemistry, either externally or with collaborators in the probe-development community. Where multiple probes existed for a target, we prioritized those in clinical development, with stronger selectivity data, or with pharmacokinetic metadata that should enable more rapid drug development. Whenever possible, concentration ranges for each compound were defined by review of literature to ensure plating at the optimal concentration for sensitivity measurements.
Cancer cell lines (CCLs) were selected from among human cancer cell lines available primarily from public organizations such as ATCC and were drawn from lineages being studied by the NCI’s TCGA program and to align with genome-wide genetic perturbation screening efforts funded by NCI (such as lung, haematopoetic/lymphoma, large intestine, and ovarian). Most CCLs have been characterized for global gene expression, gene copy number (amplifications/deletions), somatic mutations in 1,645 cancer genes, and lineage and histological subtypes (see https://portals.broadinstitute.org/ccle).
Small-Molecule Sensitivity Profiling
We profiled human CCLs in parallel in 384-well (CTRP v1) or 1536-well (CTRP v2) plates, with each cell line propagated and carefully expanded in its preferred medium. To determine the optimal density for profiling, each CCL underwent an assay-development step with a control compound (staurosporine or MG-132). Cells were plated at the optimal density, compounds were pinned in 8-point (CTRP v1) or 16-point (CTRP v2), 2-fold concentration series, and sensitivity was assayed using CellTiter-Glo, which measures cellular ATP levels as a surrogate for cell number and growth. Raw data were merged with assay metadata, and percent-viability scores were calculated relative to DMSO controls, after which concentration-response curves were fit for percent viability. The areas under percent-viability curves were computed and used as the measure of sensitivity. These data are being analyzed by our Center using a number of computational approaches to relate patterns of sensitivity to cellular and genetic features of CCLs. Connections and visualizations based on these approaches are available via the CTRP website.
Annotated Cluster Multidimensional Enrichment (ACME) Analysis
For CTRP v2 data, we developed a new data-mining approach, annotated cluster multidimensional enrichment (ACME) analysis. Unlike previous methods, which aim to identify response biomarkers with compounds individually, ACME analysis detects multiple small molecules sharing a common protein target that perform similarly across CCLs. Concurrently, ACME identifies common genetic or cellular features among multiple CCLs that share a pattern of response to small molecules, establishing a connection between the protein target and candidate response biomarkers. See Seashore-Ludlow et al. for computational details.
For CTRP v2 data, we evaluated whether small-molecule sensitivity could be correlated with basal gene-expression levels or copy-number variation to identify new biomarkers of sensitivity, candidate protein targets, or other biological determinants of small-molecule sensitivity. We performed correlation analyses both across all cancer types (“pan-can”) as well as within each lineage or subtype for which a sufficient number of CCLs was available. Pearson correlation coefficients were normalized between analyses using Fisher’s Z transformation to account for different numbers of CCLs participating in different analyses. See Rees et al. for computational details.
If you cannot access the manuscript, or if you have additional questions, please email Paul Clemons.
Access the CTD2 Data Portal.