Speakers: Duane C. Hassane (Weill Cornell Medical College), Rui-Ru Ji (Bristol-Myers Squibb), and Aravind Subramanian (Broad Institute of MIT and Harvard University)
Organizers: Andrea Califano (Columbia University), Manuel Duval (Network Therapeutics Inc.), Aris Economides (Regeneron Pharmaceuticals), Gustavo Stolovitzky (IBM Research), and Jennifer Henry (The New York Academy of Sciences)Presented by the Systems Biology Discussion Group
Reported by Kahn Rhrissorrakrai | Posted March 16, 2012
Over the past 15 years, scientists' ability to perform very large-scale screens of the transcriptomic and proteomic states of a range of organisms—from bacteria to humans—has grown tremendously. With the advent of the microarray, next-generation sequencing technologies, and mass spectrometry tools, researchers are well positioned to evaluate potential drug targets quickly. The Systems Biology Discussion Group of the New York Academy of Sciences met on October 24, 2011, for The Reactivity of the Cellular Transcriptome to Xenobiotic Compound Perturbation symposium to discuss the pharmacological uses and the discovery of small, nonessential molecules that are capable of affecting biological processes in diverse and interesting ways. In particular, these molecules affect biological processes by altering the transcriptome, the set of all RNA molecules, coding and non-coding, that are transcribed in a given cell or in a particular population of cells.
Modern genome-wide expression array (microarray) technology allows researchers to screen potential drug candidates for their effects on the cellular transcriptome and to cross-reference these effects against known transcriptomal manifestations of particular diseases. A principal challenge of drug discovery, however, is how to grapple with the massive combinatorial space that arises when researchers want to screen a library of compounds against one or more of the 40,000 human transcripts—any one of which may be differentially expressed under various cellular conditions and at various times. In addition, microarray technology, used to profile gene expression, is expensive and generally limits the range of compounds that can be tested at once. Though precise, the process of correlating entire transcriptomes with disease states is extremely time- and resource-intensive. New methods are needed to reduce costs, increase throughput, and improve the informational power of current and emerging platforms.
One approach to reducing the combinatorial burden of drug candidate testing is to narrow the field of transcripts of interest, so that every transcript need not be compared for every candidate. Aravind Subramanian, of the Broad Institute of MIT and Harvard University, discussed how this approach might be applied as part of the experimental design prior to a drug screen without sacrificing gene expression information on the full transcriptome. Subramanian and colleagues have been developing a "Connectivity Map" (also called "CMap"), a map that correlates the expression patterns of disease states (genome-wide disease signatures) with a growing list of drug candidates and with the effects of genetic perturbations (e.g., RNA interference or overexpression of genes). But, instead of correlating entire transcriptomes with perturbation, Subramanian and colleagues have identified 1000 "landmark genes" (L1000)—characteristic genes from expression clusters—that can be used to infer large portions of the transcriptome to more accurately capture the "effective dimensionality" of the transcriptome, given that genes are correlated. This system offers a highly representative expression signature from a reduced transcript set, and it is a low cost, high-throughput platform for discovering the "functional connections between drugs, genes and diseases."
These expression clusters were identified by principal component analysis (PCA) of a large expression compendium of Affymetrix (microarray) data from the Gene Expression Omnibus (GEO, a public functional genomics data repository). Subramanian's team used linear regression models to find genes that were widely expressed and displayed predictive power for dependent genes, the remaining transcripts in the cluster. While Subramanian acknowledged that this method is dependent upon the training data, he believes L1000 will perform well for a number of applications. Using these 1000 landmark genes, the system was able to recapitulate clustering results similar to those recorded when using the full transcriptome.
Landmark gene expression levels were measured with ligation-mediated amplification in the Luminex-bead system. The current system uses 384-well format with a cost of about $5 per profile and throughput of 20 plates per week. Of course, the method is not limited to drug compounds but can be extended to any molecules amenable to high-throughput dispensation, such as microRNA molecules. Subramanian's group has created a preliminary dataset of 1800 compounds, approximately 1100 RNA interference knockdowns, and 550 ORF overexpression experiments in 10 cell lines to generate a connectivity map that describes the effects of small molecules and genetic perturbations. They have been able to perform early validation of the L1000 system's ability to determine context-specific connections. For example, the team screened the anti-diabetes drug Rosiglitazone, a PPARγ (peroxisome proliferator-activated receptor γ) agonist, for its interaction with prostate cancer cells (PC3) and breast adenocarcinoma cells (MCF7). The query revealed associations with drugs known to be similar, and, more importantly, only showed associations with the PC3 cells, which—unlike MCF7—express the PPARγ receptor gene. The results from an L1000 screen can be used to quickly generate new hypotheses because when any landmark gene is implicated as a target, many of its dependent genes are also associated. Thus, this method successfully aids in rapidly reducing the number of targets that need to be screened in more detail.
Using 1000 "landmark genes" and inferring further connections from the representative clusters, Subramanian's team recapitulated many of the same connections that are seen with genome-wide microarray screening, thus demonstrating the utility and efficiency of limiting analysis to "landmark genes." (Image courtesy of Aravind Subramanian)
As important as generating a large database of pharmacological profiles is, understanding the dose response of drugs in an animal system is critical to characterizing overall efficacy. Rui-Ru Ji, of Bristol-Myers Squibb, has been developing algorithms to better understand this dose-response relationship. Many compounds hit multiple targets, e.g. a kinase inhibitor can affect multiple kinases due to domain homology. Furthermore, these targets can respond to varying levels of compounds, suggesting there is no single "correct dose." For example, Dasatinib, a small-molecule inhibitor used to treat chronic myelogenous leukemia, targets different kinases at particular concentrations ranging from 0.1 nM to 10000 nM.
To date, most methods that aim to correlate the transcript signatures associated with particular compounds do not account sufficiently for the dimensionality of microarray data (i.e., the number of microarray chips needed for each experiment) or cannot provide quantitative dose-response information. Ji presented an experimental design and computational algorithm, along with unique visualization tools, that address both challenges. The researchers set up dose-response experiments over a 6-log range with 12 doses in cell-based assays. This economical design requires only 12 arrays per compound and provides a statistically comprehensive data set without the need for point analysis. Since most drug dose responses follow a sigmoidal curve, Ji implemented a novel curve-fitting algorithm, Sigmoidal Dose-Response Search (SDRS), to be applied in genomic data sets. SDRS is a completely automated method that identifies transcripts that follow a sigmoidal dose curve over a parameter space limited to ensure realistic values. Unique to SDRS, an F-score is produced for each dose to measure the search's success in identifying probe sets, or genes, that fit the sigmoidal dose-response model.
With SDRS the researchers were able to compare the differences in potency of two compounds over the same sets of responsive genes. The F-score was then used to generate a novel visualization approach, a False Discovery Rate (FDR) heatmap in which each pixel is a dose-response–regulated gene list. Using a Fisher Test for statistical significance, they compared gene lists found at different doses and/or with different compounds to determine shared common targets or mechanisms across doses and drugs. Ji presented data that clearly showed differential dose response by cell lines (for the same drug), and the team was able find specific dose ranges and the biological pathways that the compounds may be affecting together or independently. SDRS provides an automated pipeline to better understand the effects of a drug or combination of drugs on its targets over a large dose range, thereby improving the comparison of pre-clinical data to gene expression array data.
This schematic of SDRS analysis shows how this method can be used to identify transcripts that report dose response, as well as how SDRS's F-scores can be used to generate FDR heatmaps to elucidate broad transcriptomic response and to compare compound effects across cell lines and doses.
One of the big challenges in drug discovery for cancer is the complexity of most cancers that results from the heterogeneity of their cell types. Duane C. Hassane, from Weill Cornell Medical College, spoke of the limitations of targeting only specific, primitive attributes of cancer, such as metabolic rates, when cancer cells vary significantly in these attributes. He suggested a chemical transcriptomic approach, which not only looks at individually responsive transcripts as drug targets but also identifies sets of responsive genes known to have roles in particular biological processes and determines if combinations of drugs can alter the entire program to return the system to a non-disease state.
Current therapies, which target growing and dividing cells, tend to spare stem cells, which have reduced metabolism. The survival of these cells increases the likelihood of relapse. Hassane noted that more effective treatments would target a shared property of stem cells and bulk cancer cells. One such shared property is NF-κB activity, which is associated with processes mediating cell survival and which is not a property of normal hematopoietic cells. Hassane's group looked for inhibitors of NF-κB and discovered Parthenolide (PTL), an inhibitor now in phase 1 clinical trials in the United Kingdom. Hassane's group further studied the transcripts associated with PTL activity to find interactors, genes whose products affect the same processes. Using CMap, they discovered that mTOR (mammalian target of rapamycin) inhibitors had transcriptional effects opposing those of PTL, yet when mTOR inhibitors were tested in conjunction with PTL, they enhanced the NF-κB inhibitor's effects. Hassane presented evidence that PTL in combination with Temsirolimus, an mTOR inhibitor, was an effective treatment in mice that were irradiated and injected with 2 million primary acute human myeloid leukemia cells. He subsequently found cytoprotection genes among the set of PTL-responsive genes. This result led them to ask if there was a single drug that is capable of affecting both NF-κB and cytoprotection. By using chemical transcriptomics, Hassane and colleagues were able to discover AR-42, a drug that inhibits NF-κB transcription and does not activate transcription of Nrf2, a transcription factor that decreases cytoprotection.
Chemical transcriptomics allows for the rational selection of combination therapies by identifying drugs based on a checklist of requirements. Using novel combinations of existing drugs has the added benefit of bypassing stages of the long drug testing process. Hassane's group further developed their approach by isolating normal mouse hematopoietic stem cells, transducing them with cancer genes to produce malignancy, and transplanting them into a recipient mouse. Cancer cells were then exposed to a variety of compounds before being cell-sorted. Gene signatures were obtained from microarrays. With this system, Hassane was able to find genes that are up- and down-regulated in response to different compounds and to identify combinations of compounds that normalize relevant gene expressions profiles, assuming a linear combination of factors. Using this system, it is then possible to find "if/then" conditions for drug combinations and to screen for potential cancer treatments more effectively.
Use the tab above to find multimedia from this event.
Presentations available from:
Introduction: Manuel Duval, PhD (Network Therapeutics Inc.)
Duane C. Hassane, PhD (Weill Cornell Medical College)
Rui-Ru Ji, PhD (Bristol-Myers Squibb)
Aravind Subramanian, PhD (Broad Institute of MIT and Harvard University)
RegeneronLog in or Join Now to continue