Integrating phenotypic small-molecule profiling and human genetics: the next phase in drug discovery

Abstract

Over the past decade, tremendous progress in high-throughput small-molecule screening methods has facilitated the rapid expansion of phenotype-based data. Parallel advances in genomic characterization methods have complemented these efforts by providing a growing list of annotated cell line features. Together, these developments have paved the way for feature-based identification of novel, exploitable cellular dependencies, subsequently expanding our therapeutic toolkit in cancer and other diseases. Here, we provide an overview of the evolution of phenotypic small-molecule profiling and discuss the most significant and recent profiling and analytical efforts, their impact on the field, and their clinical ramifications. We additionally provide a perspective for future developments in phenotypic profiling efforts guided by genomic science.

Introduction

There are conceptual and logistical differences between high-throughput screening (HTS) and small-molecule profiling. A conventional HTS campaign typically deals with a single phenotypic readout under one set of conditions exposed to a large number of compounds. By contrast, profiling can involve multiple types of readouts under similar conditions [1], binding profiles across many proteins [2, 3], or, classically, a single readout across many cell lines. Several years ago, we reviewed the approach of small-molecule phenotypic profiling as a means of understanding the biological effects of chemical perturbation, as well as providing a blueprint for making chemical synthetic choices in constructing screening collections [4]. In subsequent years, an increasing stream of studies focused on profiling has infused the literature (Figure 1). A landmark study in the field described small-molecule profiling of the NCI-60 collection [5], in which cell-growth measurements were made across 60 cell lines, representing multiple cancer lineages, after perturbation with a common collection of nearly 4,000 small molecules. Going a step further, the authors also profiled each cell line for protein expression of 76 common molecular targets in cancer; this combination of data was used to develop a matrix algebra-based correlation analysis of compounds with proteins. This study laid the foundation for and foreshadowed recent efforts to leverage genomic data during profiling [6-8]. Technical advances in genomic sequencing and genotyping have enabled deeper characterization of cancer cell lines across multiple types of genetic features, and high-throughput measurement of such cell lines subjected to hundreds or thousands of perturbations is now possible.

Figure 1.

Cumulative increase in the number of papers in the literature focused on profiling experiments. Results were achieved by using the query “phenotypic profiling”, followed by manual annotation of each of the 100 results for their applicability to this metric. Papers were then labeled as specifically involving small-molecule profiling. We then added the literature cited in this review that did not use the query term to the total numbers.

The use of genome-wide features in evaluating profiling results means that relationships more complex than simple single-gene associations may be uncovered. Of late, such genotype-phenotype correlations have been particularly important in cancer therapeutic discovery, as it has become increasingly clear that single drugs targeting cancer mutations may not be sufficient to treat all patients with a given tumor type. Improved diagnostics, coupled with precise small-molecule targeting, may ultimately fulfill the promise of “personalized” medicine [9-11]. Here, we review recent efforts to leverage the explosion in genomic characterization data to exploit cellular dependencies for small-molecule sensitivity and resistance. Although the cancer field has led the charge in this area, research focused on other complex genetic diseases can also benefit from the methods described here to arrive at more sophisticated models of disease and treatment.

Integrating profiling and human genetics: early efforts

Although the initial NCI-60 projects focused on determining the cancer lineage selectivity of compound response [5, 12-15], researchers discovered, through early transcriptional profiling, that histo-pathological diagnoses of tumor types did not always correspond well with the gene-expression signatures [16]. Thus, it became clear that a more refined tumor characterization would ultimately be necessary to identify correlates of small-molecule sensitivity or resistance, and a number of efforts to do so soon followed. Transcriptional profiling of the NCI-60 collection across nearly 7,000 genes, cross-analyzed with existing viability data, resulted in 232 gene expression-based signatures for compound sensitivity [17]. Again, classification was lineage-independent, demonstrating the possibility that personalized approaches may be necessary for effective treatment, as therapies tailored to tumor type were insufficient in many cases. This idea is supported by a later study, in which close characterization of response to the MEK inhibitor selumetinib showed that an 18-gene signature was sufficient to uncover activation of the MEK pathway, independent of lineage. This gene network served as a surrogate to measure MEK functional output; unexpectedly, this signature was not highly correlated with mutations in BRAF, RAS, or PI3K [18]. Together, these studies helped drive home the point that a variety of genome-wide data (e.g., gene-expression data, mutational analysis) would be necessary to uncover vulnerabilities to small-molecule treatment.

Some efforts to couple profiling to human genetics have used EBV-transformed lymphoblastoid cell lines (LCLs) derived from human B lymphocytes. Because these cell lines were used by the International HapMap Project to identify expression QTLs [19], they are densely genotyped, and thus it was thought that this cell collection could provide the rich source of genetic annotation needed for small-molecule profiling. A test of 269 such LCLs against seven compounds with distinct mechanisms of toxicity provided data to model drug response in the context of RNA transcript levels and single-nucleotide polymorphisms [20]. In this case, however, non-genetic confounders (e.g., cell growth rates, Epstein Barr virus levels used to transform the cells) prevented the study from reaching sufficient statistical power, suggesting that caution should be taken in the use of this cell source.

More disease-focused work also attempted to discover gene-small molecule interactions [21]. Using patient-derived LCLs from 18 members of a MODY1 (maturity-onset diabetes of the young 1, caused by mutation in the transcription factor HNF4a) family, the effects of nearly 4,000 bioactive compounds and clinically used drugs were tested on these cells using an ATP-based viability readout. In a similar fashion to gene-set enrichment analysis (GSEA) [22], compound-set enrichment analysis of this data revealed classes of small molecules connected with HNF4α mutation, including a subset of fatty acids. Remarkably, many of these compounds retained activity in pancreatic beta cells, enhancing glucose-stimulated insulin secretion in cells in which HNF4α had been knocked down. The authors focused on the utility of this approach in determining the functional consequence of particular disease alleles, but this work also helped lay the groundwork for profiling large numbers of genetically characterized cancer cell lines to identify genetic bases for compound sensitivity or resistance, and to uncover previously unknown cancer dependencies.

Establishing a profiling matrix

Studies in the 2000s tended to focus on one “face” of the profiling matrix (Figure 2): a particular compound class, for example, or specific types of cancer cell lines. In the former case, 500 cancer cell lines were profiled against 14 kinase inhibitors with known targets and well-annotated mechanisms of action [23]. The particular kinases targeted were examined in these cell lines for relationships between inhibitor effects and mutational status, and known relationships were confirmed, foreshadowing larger-scale efforts to come. Although many of the cellular sensitivities uncovered by this approach recapitulated known disease genotype-phenotype connections, they highlighted the existence of low-frequency, drug response-associated genetic events in tumor types not broadly considered to be drug-sensitive. These studies reinforced the notion that, at least for this particular class of compounds, the underlying genetic events are a stronger determinant of sensitivity and response than disease subtype, lineage, or tissue of origin. On the other hand, a recent effort to profile two Ewing's sarcoma cell lines against 200 kinase inhibitors [24] could be thought of as the transpose of such a matrix, where in this case only a few cell lines were profiled against many compounds. This study revealed that tozasertib possessed unique and selective efficacy against these cells; chemical proteomics identified inhibition of both Aurora A and B as responsible for this activity. This result may be the case of examining a narrow biology and small number of cell lines, however, and may not be borne out by broader experiments.

Figure 2.

The evolution of a profiling matrix over the course of the last twenty years. The three axes represent perturbations (small-molecule or genetic), cell lines or cell states profiled, and genetic features characterized. 1: a typical high-throughput screening campaign. 2: experimentation exemplified by the NCI-60 collection [5]. 3: the inclusion of gene-expression data in profiling experiments [72, 73]. 4: profiling of a small number of related compounds across many cell lines [21]. 5: the transpose of 4, representing the profiling of many compounds across a small number of related cell lines [22] (not in version 2). 6: early attempts to expand the genetic features characterized across profiled cell lines [26]. 7: efforts to expand the matrix to incorporate many cell lines, many genetic features, and now many compounds [6-8].

Alongside these more unbiased methods, approaches designed to look at particular genetic or small-molecule hypotheses were developed and applied to small-molecule profiling. For example, it had been proposed that RAS and BRAF mutations would be similarly sensitive to MEK inhibition, as mutations are often mutually exclusive, and impinge upon similar downstream signaling cascades. However, the discovery that cells harboring the BRAF mutation V600E, regardless of lineage, were selectively sensitive to MEK inhibitors over cells harboring RAS mutations [25] lent substantial credit to the promise of personalized medicine, and has led to the development of vemurafenib, a potent MEK inhibitor used for treating malignant melanoma [26, 27].

Moving towards full integration: expansion of genetic features

Previous studies successfully demonstrated that phenotypic profiling had great power to identify compounds with unique sensitivity profiles across cancer cell lines. However, it was clear that the full potential of this approach would only be realized when linked to additional cellular features. Historically, performing small-molecule screens on characterized cell line models required a series of compromises, either in cell-line, genomic-feature, or small-molecule space. Although the potential of a truly comprehensive effort was clear, the financial and logistical challenges required represented a formidable challenge. To address this challenge, a panel of 84 non-small cell lung cancer cell lines was profiled against twelve compounds under clinical evaluation [28]. A key difference in this study was the high degree of genomic characterization of the cell lines, including analyses of copy number variation and oncogenic mutation in addition to gene-expression profiling. Importantly, this additional characterization revealed the genomic similarity between cancer cell lines and primary tumors, making possible the large-scale profiling for which cell lines are required. The authors found that KRAS mutations sensitized cells to Hsp90 inhibition, whereas increased copy numbers of ABL2 and SRC induced sensitivity to the SRC/ABL inhibitor dasatinib. Such observations would likely be missed with only gene-expression profiling data.

This and other studies highlighted the need to maximize cellular models and associated cellular features (genomic or otherwise) to fully interpret the phenotypic patterns elicited by small-molecule collections. Recently, two parallel efforts from the Sanger Institute/Massachusetts General Hospital and the Broad Institute/Novartis have greatly expanded the matrix, assembling collections of 639 (Sanger) and 949 (Broad) individual cancer cell lines [6, 7]. These collections were deeply characterized for mutations in known cancer genes, whole-transcriptome mRNA expression, DNA copy number alterations, and sensitivity to a panel of small molecules (130 in the Sanger study, 24 in the Broad study). Collectively, these efforts have had a profound impact on the ability to discern genetic markers of drug sensitivity and resistance within these collections, including RAF/MEK inhibitors in melanoma and bromodomain inhibitors in N-Myc-amplified neuroblastoma [29, 30].

Cancer cell-line profiling studies, and especially the analysis of such rich datasets, are still in a developmental phase. The Sanger/MGH [6] and Broad/Novartis [7] studies used different assay technologies in their profiling experiments and subsequently used different data-analysis methods, including distinct methods of summarizing cell-line sensitivity data and differential emphases on single- versus multi-feature prediction of important genetic features. Such differences in methods, both experimental and analytical, have drawn criticism from some quarters. For example, in a meta-analysis of these two studies [31], the authors explicitly compared the concordance of both the gene-expression and pharmacological assay measurements that underlie the two studies, focusing particularly on 15 small molecules tested in both studies. The authors noted high concordance among the gene-expression measurements between the studies, but relatively low concordance between the sensitivity measurements. They interpreted these inconsistencies to mean that all of the measured pharmacological responses of both studies are questionable, as it was not clear from the available data which study provided the most meaningful assay.

A more nuanced view [32] suggested that differences between the two studies might be explained by any combination of differences in expression profiles, sensitivity measurements, or computational methods. In particular, it was pointed out that many differences in the experimental methods between the studies that might be expected to influence the results, including different assays, compound concentrations, and cell-culture conditions. Clearly, cancer cell-line profiling with small molecules requires maturation both in data-collection and - analysis procedures (relative to, say, the more mature gene-expression profiling). But it is precisely by performing these large-scale experiments that the research community will learn which methods are more likely to succeed in uncovering new cancer dependencies. As such, more attention should be focused on explaining discrepancies between these studies in terms of the real, technical differences between them.

Similar to the model advanced by the NCI through the NCI-60 collection, these datasets are not static: they are modular entities that can be expanded in the future (see Figure 2). For example, the Broad/Novartis Cancer Cell Line Encyclopedia (CCLE) provides deep genomic characterization of the ∼950 cell lines in the collection, but was tested against relatively few (24) compounds. It was clear that a much larger collection of small molecules needed to be profiled to maximize the power of this approach. To that end, Basu, Bodycombe, Cheah et. al. acquired a subset of 242 CCLE cancer cell line models and profiled them against an expanded collection of 354 small molecules, selected for their high selectivity and targeting of many proteins known or thought to be hubs of cell signaling [8]. In addition to expanding the overarching chemical-biology coverage of this dataset, this study highlighted an unappreciated dependency on anti-apoptotic Bcl-2 family members in the context of activating β-catenin mutations. Collectively, these efforts suggest that our discovery capabilities are not saturated, and that enhanced cell models and compound collections are likely to further our ability to identify new therapeutic hypotheses and to refine those nominated in earlier studies.

Importantly, Basu, Bodycombe, Cheah et al. [8] also explicitly exploited cell lineage or tissue of origin as a variable in their analysis, and highlighted the importance of confounding factors such as whether cells were grown in suspension or were adherent to the tissue culture substrate. This study focused primarily on the association of mutation and lineage features of CCLs with response to small molecules, resulting in a large number of such connections and a public resource which users can query using compounds or genes of interest [8] (http://broadinstitute.org/ctrp). A more recent study from the same group systematically examined correlations of small-molecule sensitivity for 481 compounds with genome-wide basal gene expression among 860 CCLs. Such analyses revealed unique and diverse insights into drug metabolism and mechanism of action in cancer cells, including identification of new target pathways and partner proteins (Rees et al., manuscript submitted). Although the sensitivity measurements used were only correlated with gene-expression data, by amassing enough data across a large number of cell lines, the authors were able to discern significant patterns of connecting expression to compound sensitivity.

Using multi-gene features to find genomic correlates of chemical sensitivity

With the advent of large pharmacological datasets, finding genomic correlates of drug sensitivity will increasingly need to leverage the predictive power of using sets of multiple genomic features. The discoveries made using single-gene approaches likely represent only the tip of the iceberg, and we are far from saturation. The most obvious multi-gene approaches rely heavily on the body of existing knowledge, such as pathway databases and protein-protein interaction networks. Methods such as GSEA [22] offer a starting point, though their utility in analyzing these experiments depends heavily on the choice of gene-set collection used, which must be specified in advance [33]. One study proposes an alternative gene set collection, based on co-expressed genes in cancer, to those available in GSEA as superior for cancer cell line profiling data, as more of the gene sets in this collection were significantly enriched after treatment of cancer cell lines with a collection of small molecules [33]. Indeed, many methods rely on the state and completeness of sources of prior knowledge, as such data may be incomplete. The relevance of such prior knowledge is itself a subject of study, in which network biology models having or lacking such prior information are directly compared [34]. Unbiased methods, such as elastic net and related regression techniques [35], have been used to discover cancer dependencies [7, 8], but their application to these data had not been systematically evaluated until recently [36]. In general, models emerging from multi-gene-predictive methods still require biological interpretation of the lists of genes that represent model features. Most such analyses tend to center on genes recognizable by name to cancer biologists.

One way to analyze cancer cell-line profiling to exploit dependencies based on multiple features is to create a signature of a particular cell state (e.g., glycolytic versus respiring [37, 38]) and correlate this signature with sensitivity to small molecules. Cell-state signatures are in principle accessible using any ‘omic data, but most commonly from gene-expression data. In a recent example, Viswanathan et al. (manuscript submitted) used gene-expression and pathway analysis to prepare a signature of cell lines’ progression relative to the epithelial-to-mesenchymal transition (EMT). By correlating this signature with small-molecule sensitivity data, these authors showed that inhibitors of glutathione peroxidase 4 (GPX4) exhibit selectivity for epithelium-derived cancer cells that have undergone EMT, as well as for cancer cells of mesenchymal origin.

Another recent study explored the associations between groups of small molecules targeting common proteins or pathways and groups of cell lines sharing cellular or genetic features (Seashore-Ludlow et al., manuscript submitted). Clustering sensitivity data, followed by mining clusters for enriched hot-spots, allows unbiased identification of clusters of CCLs sharing common genetic features that respond similarly to small molecules with shared targets. These studies provide a blueprint for moving toward profiling analyses that incorporate a greater proportion of overall genomic data to identify correlates of small-molecule sensitivity.

Identifying resistance mechanisms through profiling

The identification of small-molecule dependencies in cancer has had a direct and profound impact on clinical outcomes. Despite this progress, the long-term efficacy of anti-cancer therapies, including both targeted and cytotoxic agents, is almost universally thwarted by the acquisition of drug resistance, representing an unmet clinical challenge. Moreover, as our collection of therapeutic modalities expands, this challenge is likely to continue to grow in scope and size, increasing the necessity for understanding the genetic underpinnings of therapeutic resistance. For some time, drug resistance appeared to be a complex and insurmountable phenomenon. However, the advent of systematic methods of identifying resistance mechanisms has, similar to pharmacogenomic screening approaches, permitted the community to start defining the principles of therapeutic resistance. Much of this progress has leveraged the development of comprehensive genetic perturbation libraries, including gain- (cDNA/ORF) and loss-of-function (shRNA/CRISPR) reagents, but the principles followed are the same for small-molecule profiling. These reagents have successfully been employed as a systematic and comprehensive means of rescuing the phenotype associated with small-molecule dependencies. An initial demonstration of this approach used a genome-scale shRNA library to identify loss-of-function events that drive therapeutic resistance to imatinib in a cell line model of CML, leading to identification of PTPN1, NF1, SMARCB1, and SMARCE1 as genes essential for imatinib response [39].

More recently, cDNA (or open reading frame/ORF) gene-expression libraries have caught up with loss-of-function libraries, enabling screening for genes sufficient to induce therapeutic resistance [40]. The initial iteration of this experiment involved the screening of a kinase-centric library, which was successfully used to identify COT/MAP3K8 as a driver of resistance to RAF inhibitors in BRAFV600E-mutant melanoma. Subsequent efforts employing this screening resource identified a number of kinase mediators of resistance, including p90RSK3/4 (PI3K inhibition [41]), CRKL (EGFR inhibition [42]) and PKA (HER2 inhibition [43]). Recently, this approach has been expanded beyond the kinome to a near-genome scale collection of cDNAs. Using malignant melanoma as a model system for understanding therapeutic resistance, Johannessen et al. [44] screened this collection for genes that mediate resistance to RAF, MEK, ERK, and combined RAF/MEK inhibition in the context of the BRAF^V600E mutation. These studies revealed a complex and diverse set of kinase and non-kinase resistance effectors, including transcription factors, signaling pathways operant outside of the RAF/MEK/ERK pathway, and a series of other uncharacterized genes. In addition to loss-of-function and wild-type ORF libraries, mutational scanning, for example of kinase drug targets [45-48], has been a particularly useful tool for identifying resistance mutations. Recent technologies enabling the creation of fully-saturated mutational scanning libraries are likely to further increase our understanding of drug-target interactions, as well as annotating resistance alleles [49]. Together, these studies advance the notion that large-scale profiling experiments can uncover genetic targets for therapeutic intervention.

Outlook for the future

A key feature enabling the success of a genomic approach to phenotypic drug discovery has been the detailed level of characterization of cancer cell lines, and the fact that such models have been shown to faithfully recapitulate the genetics and small-molecule sensitivity of primary patient samples [25, 50]. In order to expand beyond cancer, however, disease-focused communities need to create similar tools and models to enable modernization of drug discovery in their disease areas. A recent example is provided by a chemical genomics approach taken toward malaria [51]. In that case, 2,816 compounds were profiled in eight-point concentration-response against 61 parasite lines, and the results were combined with genome-wide association studies and linkage analysis for selective sensitivity. In particular, compounds with differential chemical phenotypes across three parental parasite strains were investigated more fully. Remarkably, 96% of these differences were mapped to three loci containing genes encoding dihydrofolate reductase, multidrug resistance protein, and chloroquine resistance transporter. One outcome resulting from such a dataset is a set of novel strategies for overcoming drug resistance in malaria, but the point that stands out here is the high degree of analogy to cancer cell-line profiling. In both cases, the goal is to find genetic determinants of susceptibility to small-molecule perturbation. Further, the similarity between resistance in cancer and infectious disease has been recognized [52].

Exploring beyond the DNA-encoded genome for cellular vulnerabilities to small-molecule perturbation is likely to be an important next direction for phenotypic profiling. For example, epigenetic modifications represent reversible changes that might be leveraged in designing profiling experiments [53], and a number of chromatin-modifying enzymes are under heavy scrutiny, such as EZH2 and DOT1L [54-56]. Similarly, as proteomics and metabolomics techniques become more miniaturized and routine, profiling the functional genome [57, 58] or metabolome [59, 60] for cancer cell dependencies will become more of a reality. Analysis methods, such as similarity network fusion [61], optimized to take advantage of such integrated datasets, are already in development. Such methods will be particularly important as we move toward larger datasets, for example those that employ genomic measurements following each perturbation, as opposed to in the steady state.

The development of disease-relevant readouts and models is absolutely essential for the success of linking genotype to phenotype. Again, this approach has been most evident in the cancer community, where the NCI-60 collection has had so much influence on the field [5, 15]. More recently, deliberate efforts to create panels of cell lines have enabled greater statistical power across lineages and mutational status [62]. Lymphoblastoid cell lines, for example, derived from the International HapMap project [63], may have their own shortcomings regarding reproducibility and other confounding factors [20]. Ultimately, the use of techniques to generate induced pluripotent stem (iPS) cells from patient samples, followed by directed differentiation to relevant cell types [64-66], may be the most productive avenue for the future. It remains to be seen if these approaches can bear fruit in cases where cell death is not the desired phenotype, but the establishment of an NCI-60-like collection for other diseases would go a long way toward answering this question.

Table 1. Table of phenotypic profiling efforts.

Year	Area of focus	Cell lines	Compounds	Ref(s)
1988-current	growth inhibition	60	ongoing	5, 12-15
2008	quantitative trait loci for compound sensitivity	269 LCLs	7	20
2011	MODY1	18 LCLs	∼4000	21
2007	kinase inhibitor profiling	500	14	23
2011	kinase inhibitor profiling in Ewing's sarcoma	2	200	24
2009	genomic characterization of non-small cell lung cancer	84	12	28
2012	broad cancer cell line characterization	639	130	6
2012	broad cancer cell line characterization	949	24	7
2013	expansion of small molecules tested	242	354	8

Text Box. Recent advances in small-molecule profiling.

A method of small-molecule profiling that has begun to be explored is building profiles using historical screening data. Such profiling is most useful for screening collections exposed together to many assays over time, usually in the same facility, so that a large number of measurements accumulates. Even though the assays were individually conceived and not intended as a profile, it is possible to normalize or discretize such data to make meaningful calculations of performance similarity between compounds. In one recent study [67], computational scientists profiled historical screening data across 195 assays run on ∼1.5 million compounds. The authors show applications for their method in developing hypotheses about mechanism of action (MoA), lead-hopping, and selection of library subsets. A similar study profiled historical screening data [68] and likewise showed the ability of such data to generate MoA hypotheses and to group compounds into communities based solely on their performance. Whenever a common compound collection is exposed to many such assays, these methods hold great promise, a fact recognized by the NIH when it commissioned a new public interface, the BioAssay Research Database (BARD), to showcase the findings of the Molecular Libraries Program [69] (Howe et al., manuscript submitted).

Though powerful, methods that require accumulation of data are inherently passive: they require that datasets be allowed to accumulate over extended periods of time. In each of the above cases, approximately ten years of data collection were performed before the data-mining methods were published and presented. A more active approach to profiling is to expose small-molecule collections prospectively to high-dimensional multiplex measurements; recent advances in assay technology make this possible. High-content screening, usually involving high-throughput microscopy and image analysis, has often been at the forefront of method development [70-72]. A recently developed imaging assay using multiple fluorescent dyes and channels was used for unbiased cell-morphological profiling of small-molecule action [73], analogous to the more familiar gene-expression profiling of small molecules [74, 75]. In general, high-dimensional data collection per well and hierarchical clustering methods have been used to study MoA for novel compounds [75, 76], and can also be used to develop profile-based structure-activity relationships among molecules [77]. In one recent study, a large collection of compounds was exposed to profiling by both multiplex imaging and gene-expression methods, and the resulting data used to select a performance-diverse subset of compounds [76]. These technological advances reduce some of the barriers to assembling informer sets for small-molecule profiling, freeing up time and effort to devote to more sophisticated data analysis and integration.

Highlights.

Integrating phenotypic small-molecule profiling and human genetics: the next phase in drug discovery Cory M. Johannessen, Paul A. Clemons, Bridget K. Wagner

• Drug and target discovery will benefit from cell-line genomic characterization

• Quality cell models and selective small molecules are critical tools for success

• The cancer field has used this approach, but other diseases would benefit, too

Glossary

Bioactive compounds

small molecule perturbagens with previously known and annotated mechanism(s) of action

Cancer Cell Line Encyclopedia (CCLE)

A project led by the Broad Institute, the Novartis Institutes for Biomedical Research, and the Genomics Institute of the Novartis Research Foundation, focused on the genomic and pharmacologic characterization of a set of ∼1000 human cancer cell lines. http://www.broadinstitute.org/ccle/home

Cellular dependency

pathway or process on which a particular cell type or cell state depends for survival

Chemical proteomics

measurement of the protein(s) to which a small molecule binds, directly or indirectly

Gene-set enrichment analysis (GSEA)

A computational method to determine the enrichment of a defined set of genes in the gene-expression analysis of two biological states. The collection of gene sets under consideration can be derived experimentally, computationally, or by biological function.

Interaction network

a graph-theoretic construct that expresses relationships between proteins based on evidence of interaction between them (may be physical interaction, co-regulation, or combinations of evidence)

Lead-hopping

The identification of novel chemical matter with similar biological function, but different structure, to a starting small molecule.

Lymphoblastoid cell line (LCL)

Peripheral B lymphocytes transformed by Epstein-Barr virus, these cells are often useful as a source of DNA and cells from individuals for biological characterization.

Matrix algebra

mathematical technique allowing connection of two variables via a third shared variable that connects to each one

Meta-analysis

an analysis technique that merges data from separate studies of the same subject(s) into a single analysis

Metabolomics

direct measurement of the levels of small-molecule metabolites in a cellular system under various conditions

Molecular Libraries Program

A program operated by the National Institutes of Health focused on the discovery of novel chemical probes of biological processes and disease states. http://mli.nih.gov/mli/mlp-overview/

Multi-feature prediction

mathematical technique that combines information from multiple predictive variables to estimate the value of a predicted variable

NCI-60 collection

A long-standing project operated by the National Cancer Institute involving the screening of a collection of 60 cancer cell lines for growth inhibition after treatment with small molecules of interest. http://dtp.nci.nih.gov/branches/btb/ivclsp.html

Performance similarity

using performance of small molecules across multiple assays or using multiple features as a measure of their similarity (instead of, for example, chemical structure similarity)

Phenotypic pattern

a quantitative measurement of phenotype spanning multiple variables, such as gene-expression in response to compound treatment

Quantitative trait locus

a genetic determinant of a quantitative trait of an organism, such as size

Signature

specific set of multiple features reporting on a disease state or action of a perturbagen

Small-molecule profiling

unbiased measurement of multiple cellular features reporting on small-molecule action

Transcriptional profiling

unbiased measurement of multiple gene-expression features (usually as transcripts, possibly genome-wide)