Genome-Wide RNAi Analysis of Growth and Viability in Drosophila Cells

See allHide authors and affiliations

Science  06 Feb 2004:
Vol. 303, Issue 5659, pp. 832-835
DOI: 10.1126/science.1091266


A crucial aim upon completion of whole genome sequences is the functional analysis of all predicted genes. We have applied a high-throughput RNA-interference (RNAi) screen of 19,470 double-stranded (ds) RNAs in cultured cells to characterize the function of nearly all (91%) predicted Drosophila genes in cell growth and viability. We found 438 dsRNAs that identified essential genes, among which 80% lacked mutant alleles. A quantitative assay of cell number was applied to identify genes of known and uncharacterized functions. In particular, we demonstrate a role for the homolog of a mammalian acute myeloid leukemia gene (AML1) in cell survival. Such a systematic screen for cell phenotypes, such as cell viability, can thus be effective in characterizing functionally related genes on a genome-wide scale.

The availability of entire genome sequences stimulates the advancement of functional genomic approaches, both to accelerate the comprehensive identification of components in biological programs and to understand their functional conservation across species (1, 2). In yeast and in worms, genome-wide phenotypic studies identified genes that were essential for cell fitness and development (3, 4). Drosophila is one of the best-studied genetic organisms and is instrumental to the identification of conserved pathways with important roles from flies to humans. Treatment of cultured Drosophila cells with dsRNA leads to the depletion of the corresponding transcript and the generation of specific and penetrant phenotypes (5), providing an efficient approach for systematic loss-of-function phenotypic analyses (69).

Cell growth, proliferation, and survival are fundamental processes that maintain cell populations and impinge on lineage expansion and pattern formation (10). To identify gene functions by cell-based RNAi screens, we generated a dsRNA library targeting nearly all genes in the Drosophila genome (fig. S1). In total, 21,306 primer pairs were designed to amplify gene-specific fragments (11) then used for synthesis of dsRNAs (12). We established a quantitative assay of cell number that correlated the reduction of signal to dying cells, as demonstrated by RNAi of the D-IAP1 (13) inhibitor of apoptosis (a positive control in screen and data processing; Fig. 1, A and B).

Fig. 1.

Genome-wide RNAi screens result in highly reproducible growth and viability defects. Results for Kc167 cells are shown; similar results were observed for S2R+ cells (22) (table S1). (A) Left panel: Luciferase activity (relative light units) indicative of ATP levels is correlated with number of Drosophila cells in a high-throughput assay format (fig. S1). Right panel: Treatment with dsRNA targeting D-IAP1 induced time-dependent decrease in cell viability is shown as the relative readout as compared with cells treated with control green fluorescent protein (gfp) dsRNA (normalized to 1). (B) Fluorescence microscopy of cells after 3 days RNAi. More dying cells were detected after treatment with D-IAP1 than with control dsRNAs by the ratio of SYTOX green–labeled nuclei (green and lower panel) versus Hoechst 33342–labeled nuclei (red). (C) Results from one genome-wide RNAi screen, after 5 days dsRNA treatment. Each RNAi experiment is represented by a shaded box (a single well), arranged by 384-well plates as outlined in upper left. Results in each plate were mean-centered before overall analysis. Gray values indicate z score, with darker shades representing below-average results. Each 384-well plate had four control wells containing either D-IAP1 or the negative controls gfp, Rho1, or no dsRNAs. The D-IAP1 control phenotypes are evident as the dark boxes in the upper left corner of each plate, indicative of dying cells and a lower signal. (D) Example of highly reproducible phenotypes with similar z scores from two independent RNAi screens [enlarged from (C) and from duplicate screen in fig. S2].

Genome-wide RNAi screens in two embryonic hemocyte (blood cell) lines (14, 15) were each performed in duplicate (Fig. 1C and fig. S2). In total, we analyzed 77,880 RNAi experiments (12). A comparison between duplicate screens revealed qualitatively and highly quantitatively reproducible results (Fig. 1D) with a correlation coefficient of 0.86 (Kc167 cells, fig. S2). The results included reproducibly subtle to severe phenotypes as quantified by a significance factor, or z score (this score signifies the severity or rank of specific RNAi phenotypes) (12) (Fig. 1, C and D). For example, D-IAP1 resulted in a severe phenotype with z scores of 6.1 and 7.0 in S2R+ and Kc167 cells, respectively. To determine the efficiency of the screen, we assessed the phenotypes for a highly represented functional group of genes. Nearly all of the tested genes encoding ribosomal components exhibited similar quantitative phenotypes (72 genes with a mean z score of 2.9 ± 0.8; fig. S3), of which 68 had at least one z score greater than 2.2, and 54 greater than 3.0. The phenotypic scores for the total set of ribosomal components were statistically distinguishable from the cell death phenotype of D-IAP1 (P < 0.0001, Fig. 2), showing that differences in z scores between selected gene sets can be detected by this method. Quantitative reproducibility of the method was also demonstrated through an analysis of the phenotypes from independent dsRNAs targeting the same gene within one screen. For example, each dsRNA directed against neighboring predicted genes cpo, CG12349, and CG18435 exhibited similar phenotypes (z scores of 3.6, 3.2, and 3.7, respectively). Consistent with this, current annotation now depicts these three gene models as a single gene (16).

Fig. 2.

Similar quantitative RNAi phenotypes of genes encoding ribosomal proteins. Averaged RNAi phenotypes of 72 genes encoding all annotated ribosomal proteins tested (gray bars) are distinguishable from negative controls (white bars, gfp dsRNA) and more severe phenotypes (black bars, D-IAP1 dsRNA). The gfp and D-IAP1 results represent negative and positive control experiments (scored one per plate) over a genome-wide screen. Intergroup comparisons are highly significant in a Student's t test (P < 0.0001) (fig. S3).

To further analyze essential gene functions, we selected 438 cases in which dsRNAs resulted in greatly reduced cell number as represented by a z score of 3 or more (Fig. 3A and tables S1 and S2). This threshold included RNAi phenotypes ranging in severity from z scores of 3 to 7.7, likely representing defects in cell growth to defects in cell survival. Only 20% (87 out of 438) of the identified genes had associated mutant alleles in Drosophila, which included genes with demonstrated roles in cell growth, cell cycle, and anti-apoptotic cell survival (12) (table S3), whereas the vast majority lacked mutant alleles for functional analyses.

Fig. 3.

Quantitative grouping of RNAi phenotypes. (A) Distribution of the frequency of RNAi phenotypes recovered for each specified range of z scores. We used a z score of three or more standard deviations from the mean as a threshold to select 438 results for further analysis (tables S1 and S10). (B and C) Frequency of encoded functional groups as predicted by InterPro protein domains and manually assigned to representative categories (tables S1 and S4) for all selected phenotypes [(B), z score > 3] and the most severe phenotypes [(C), z score > 5], revealing significant changes in the abundance of predicted ribosome proteins (P < 0.001) and proteins with no predicted domains (P < 0.00001). z scores were averaged across experiments. (D) Classification of quantitative RNAi phenotypes of selected genes (rows) identifies groups of related and newgene functions, as determined from duplicate screens per cell type (columns) and visualized by z score (scale, bottom). Both D-IAP1* (added control) and D-IAP1 (within RNAi library) yield equivalent phenotypes.

Of the 438 cases, 47% (206 out of 438) had an associated Gene Ontology annotation (17), and 59% (260 out of 438) encoded an identifiable InterPro protein domain (18). When the most abundant domain predictions were used to categorize genes into distinct functional classes (Fig. 3B; individual predictions and assignments in tables S1 and S4) (12), the relative distribution of predicted gene functions differed with the quantitative severity of the RNAi phenotypes (Fig. 3, B and C).

The phenotypic screen also identified genes encoding sets of proteins in known biochemical complexes, as revealed by sequence-based classification (Fig. 3B). Examples of this were within two of the most abundant categories pertaining to protein translation (56 genes, “Ribosome”) and ubiquitylation and protein degradation (34 genes, “Proteasome”). Genes with predicted roles in the cell cycle showed quantitative phenotypes similar to those involved in protein translation, but in only one of the two cell types screened (“Cell Cycle,” Fig. 3D; supporting online material text). One of the most populated categories consisted of genes for 62 proteins with predicted DNA binding domains (“DNA binding,” Fig. 3B, table S1), including chromatin-related factors (e.g., core Histone and high-mobility group (HMG)–box domains: bss and CG17836; Fig. 3D) and representative members of transcription factor families (e.g., homeobox, ets, and AML domains: abd-A, aop, and CG15455; Fig. 3D). Only genes for specific transcription factors from within different families were identified. For example, although four different AML-like genes are encoded within the fly genome, only one, CG15455, was functionally identified in the screen (fig. S4). Serpent, a GATA-1 homolog with roles in fly and mammalian blood cell development and survival (19), was identified as the only one of five predicted GATA-type Zinc-finger transcription factors (srp, Fig. 3D). Proteins with a predicted DNA binding domain comprised the largest assigned category of genes identified, both in total (14%, Fig. 3B) and in the class with the most severe phenotypes (19%, z score > 5, Fig. 3C), an enrichment from the proportion found in the genome (5%).

Overall, the largest category of genes (41%) had no recognizable predicted protein domain (178 genes, “No Prediction,” Fig. 3B), suggesting that the screen identified many uncharacterized genes with essential cellular roles (table S1). For example, severe cell viability phenotypes (z scores 6.8 and 7.3) were observed with HFA13298 dsRNA targeted against a newly predicted gene, HDC14318, with six overlapping expressed-sequence tags mapped to the same region (RH13972, RH23223, RH26651, RH22174, RH26647, and RH62785). The proportion of 438 genes with phenotypes but without a predicted protein domain increased with phenotypic severity (63% “No Prediction,” z score > 5; Fig. 3C).

We also identified uncharacterized genes with phenotypes quantitatively similar to that of D-IAP1 (z score > 5), raising the possibility that these loss-of-function phenotypes resulted from cell death, perhaps due to the activation of apoptosis. We further evaluated two such genes, CG11700, a ubiquitin-like gene, and CG15455, a gene encoding an AML1-like transcription factor (z scores 7.2 and 7.4 in Kc167 cells, respectively, Fig. 3D). The phenotypic severity could not be attributed to an accumulated arrest in transition at one stage in the cell cycle (12) (Fig. 4A and fig. S5). As indicated by terminal transferase-labeled DNA breaks, over 95% of cells treated with dsRNA to CG11700 or D-IAP1, and 20% of cells treated with dsRNA to CG15455, were apoptotic (12) (Fig. 4B). The addition of a pan-caspase inhibitor, Z-Val-Ala-DL-Asp(O-Methyl)-fluoro-methylketone (z-VAD-fmk), reverted the cell death in response to the RNAi of CG11700 and D-IAP1, and to a lesser extent of CG15455 and other transcription factors (Fig. 4C). D-IAP1 directly inhibits the proapoptotic caspase, Nc (Nedd2-like) (20). The CG11700 and D-IAP1 dsRNA-induced cell death phenotypes were both rescued by combined RNAi removing the single Nc caspase function (Fig. 4C). In contrast, neither the loss of function of Nc (Fig. 4) nor the loss of function of the transcriptionally activated proapoptotic gene reaper (19, 21, 22) (fig. S6) was sufficient to suppress cell death upon co-RNAi with CG15455 or other tested transcription factors. Together, these results suggest that the ubiquitin-like CG11700 may act in the same pathway as D-IAP1 to directly prevent Nc caspase-activated apoptotic cell death. In contrast, a set of essential transcription factors may regulate complex responses for cell fate, proliferation, and/or cell survival that directly or indirectly initiate a partially caspase-dependent apoptotic program.

Fig. 4.

Different anti-apoptotic gene functions identified by severe RNAi viability phenotypes. Experiments shown as assayed in Kc167 cells following RNAi against a negative control (gfp) and D-IAP1, CG11700 (ubiquitin-like), and CG15455 (AML-1-like, fig. S4) genes, each identified in the screen by severe phenotypes (z scores 7.0, 7.2, and 7.4, respectively). (A) Flow cytometry analysis of propidium iodide (PI) stained DNA after 3 days RNAi, as indicated (12). Analyses of total events reveal decreased cell size and DNA content, indicative of dying cells. Cell cycle distribution analysis performed on viable cells with ≥2N DNA content [percentage of total events shown, fig. S5 (12)]. FSC, forward scatter channel. (B) Two classes of severe RNAi phenotypes were distinguished by the proportion of apoptotic cells (>95 and 20%), as indicated by fluorescein-labeled dsDNA breaks [terminal deoxynucleotidyl transferase–mediated deoxyuridine triphosphate nick end labeling (TUNEL), green and lower panel] versus total cell nuclei (Hoechst 33342 DNA stain, red) 7 days after treatment with dsRNAs. (C) Left panel: Rescue of RNAi growth and viability phenotypes by a pan-caspase inhibitor (z-VAD-fmk). Shown are the ratios between combined treatments with dsRNA and either z-VAD-fmk in dimethyl sulfoxide (DMSO), or in DMSO alone (red line, normalized to 1), from results of averaged triplicate experiments. Right panel: Data are displayed as in the left panel, but they are from combined treatments with test dsRNA and either dsRNA against Nc caspase or gfp control.

In comparisons made between complete proteomes (12), the percentage of predicted orthologs found for the genes with RNAi viability phenotypes was higher than the percentage of orthologs found in searches of the entire Drosophila proteome with those from yeast, worm, mosquito, mouse, and human (fig. S6 and table S5). Notably, 50 genes had homology to human disease genes (table S6), including 10 genes implicated in blood-cell leukemia (e.g., AML1) and genes described with anti-apoptotic functions (FOXOA1 and MLK). Thus, functional analysis in Drosophila cells uncovered common key regulators for animal cell survival and proliferative decisions. Interestingly, in contrast to the total results, the most severe RNAi phenotypes (z score > 5, fig. S8) identified significantly fewer yeast homologs (from 39 to 19.3%, respectively), a similar percentage of animal-specific homologs (27.6 and 29.8%), and an increased number of genes without high-scoring matches (from 33.3 to 50.9%) (12). This suggests that metazoans may have evolved specific mechanisms, such as the preservation of cell identity by a specific code of transcription factors, to maintain cell viability.

Functional analysis by RNAi reveals previously unknown and evolutionarily conserved gene functions, with the powerful ability to comprehensively and quantitatively determine the contribution of potentially every gene to a particular process. Quantitative cell-based analysis offers advantages by permitting the detection of gene functions associated with subtle or redundant phenotypes in organisms. This approach also holds the potential for statistical clustering across many different cellular phenotypes to elucidate complex gene functions as more data accumulate (23). The described genome-wide RNAi library is adaptable for screening for many different cellular pathways and processes, ultimately leading to a functional understanding of cellular systems that control development and disease.

Supporting Online Material

Materials and Methods

SOM Text

Figs. S1 to S8

Tables S1 to S8

References and Notes

View Abstract

Navigate This Article