A pipeline for mapping of epistasis in mammalian cells

Our platform for the generation and quantification of GIs in mammalian cells (Fig. 1) is based on RNAi-mediated depletion of gene function in a pairwise fashion and measurement of the resulting phenotypic consequences. The availability of genome-wide RNAi libraries (small interfering RNA (siRNA), small hairpin RNA (shRNA) and endonuclease-prepared siRNA (esiRNA)) for higher organisms has enabled high-throughput genetic screens to be routinely performed in an arrayed or pooled setting¹³. Pooled approaches are based on monitoring the enrichment or depletion of specific sequence barcodes from a starting pool as a proxy for cell fitness¹⁴. In this study, we developed a platform in which the phenotypic effects of pairwise knockdowns can be quantified in a systematic manner. The measurement of every combination in the array is central in the subsequent data analysis.

esiRNA are siRNAs generated by enzymatic cleavage of a long double-stranded RNA (dsRNA), resulting in a heterogeneous mixture of siRNA sequences that all target the same mRNA sequence¹⁵. This reagent has several properties that make it preferable over other approaches that rely on chemically synthesized siRNA and shRNA. Because any single siRNA sequence is present at an extremely low concentration, esiRNAs exhibit less off-target effects (and noise) than commonly associated with single siRNAs and shRNAs¹⁶. In addition, the number of pairs to test (and hence the amount of reagent required) grows quadratically with the number of genes to be interrogated, rendering the cost of siRNA or shRNA reagents prohibitive in an arrayed format. Conversely, large quantities of esiRNAs can be produced inexpensively in a high-throughput manner (Online Methods).

In our study, we used cell count as a measure of cell proliferation capacity or fitness. A conceptually similar readout (such as colony size) has been used in yeast and bacteria^5,7,17. Data from these simpler systems suggests that strong GIs are rare^10,12; therefore a general phenotype such as proliferation capacity, influenced by many different genes and pathways, is likely to uncover GIs. Finally, this phenotype is relatively easy to monitor with unsophisticated and affordable instrumentation.

To generate E-MAPs in mammalian cells, we dispensed unique esiRNAs into a 96-well or 384-well microplate (the array), and, using a custom-built liquid handler (Supplementary Fig. 1), a single constant esiRNA was overlaid (the query) to produce all pairwise combinations with the given query (Fig. 1). We transfected mouse fibroblast cells engineered to express a nuclear marker (histone macroH2A-EGFP), and allowed them to proliferate for 72 h. We assessed cellular proliferation by counting the number of cells containing fluorescent nuclei in each well using high-throughput fluorescence cytometry with an Acumen eX3 plate cytometer (TTP Labtech). We chose the microplate format over a cell spot microarray–based method¹⁸ as it allows many more cells to be monitored (thousands rather than dozens), thus allowing for the detection of more subtle phenotypes. We tailored protocols that allowed us to perform high-throughput knockdown experiments with a screening Z factor¹⁹ of 0.57 and >90% knockdown of EGFP as positive control (Supplementary Fig. 2a). As confirmation that GIs are indeed rare, addition of the query did not on average result in large deviations from the number of cells expected from single esiRNAs alone (Supplementary Fig. 2b), and we clearly observed fitness defects associated with particular queries when combined with the esiRNA array (Supplementary Fig. 2c).

We scored raw data using a GI score (termed S score) based on a neutral interaction model centered at zero (a neutral gene pair) and developed to capture a continuous spectrum of phenotype strengths²⁰ (Fig. 1). We processed the data using a similar statistical framework similar to the one developed for yeast and bacterial GI mapping (Online Methods; http://sourceforge.net/projects/emap-toolbox/)^17,20. A different GI scoring system, termed the π score, based on similar assumptions as the S score had been developed for analysis of RNAi-based GI data derived from D. melanogaster⁹. Using our mammalian data, we found that GIs obtained using the two methods were very similar (Supplementary Fig. 3a). However, based on our data, the S score was slightly more reproducible (Fig. 2a and Supplementary Fig. 3b), allowed for rational selection of score cutoffs based on a neutral interaction model and facilitated comparisons to other organisms for which similar data have been obtained, and therefore we used it to define our GIs.

A chromatin-centered E-MAP in mouse fibroblasts

We used our epistasis-mapping pipeline to genetically interrogate factors involved in chromatin regulation in mice. We targeted 130 genes involved in chromatin or chromosome regulation using pairwise combinations of esiRNAs in mouse fibroblast cells (Online Methods). A common concern when using RNAi is that only partial phenotypes may be observed from incomplete knockdowns. However, hypomorphic (DAmP) alleles, which reduce protein levels of essential genes in yeast^12,21, frequently yield rich and biologically meaningful GI profiles, suggesting that partial gene-product depletion produced by RNAi can be informative. We verified the efficiency of mRNA knockdown and determined that the transcripts for ~60% of the selected genes were inhibited by more than 60% (Fig. 2b and Supplementary Table 1). We found that knockdown efficiencies correlated with expression levels of the target in this cell line (Fig. 2c). Furthermore, we observed that efficient knockdown of individual genes occurred when two different esiRNAs were present (Supplementary Figs. 4 and 5). We generated a data set covering ~11,000 pairwise combinations (Supplementary Data 1 and 2) and the final data is of high quality as judged by internal consistency based on correlating S scores derived from independent biological replicates (Fig. 2a). Using an empirical cutoff for the S score, which reflects a high confidence of bona fide GIs²⁰ (Supplementary Fig. 6), we uncovered 929 positive (S ≥ 2) and 611 negative (S ≤ −2) GIs (Supplementary Table 2). Finally, to assess potential off-target effects and help address the overall quality of the data set, we generated independent esiRNAs for five of the genes on the E-MAP (Cnot1, Cnot2, Cnot3, Paf1 and Leo1; Supplementary Table 3) and found that the genetic profiles they generated were highly correlated with the profiles from the original esiRNA constructs (ranging from r = 0.61 to r = 0.84; Supplementary Fig. 7). The data from these experiments were averaged and integrated into the final data set (Supplementary Data 2).

Comparison of mammalian GI and PPI data

Epistasis mapping is a powerful tool for unbiased discovery of protein complexes, larger functional modules and pathways. Data from simpler systems, including S. cerevisiae and S. pombe, suggest that genes in functional modules exhibit two important characteristics: (i) their phenotypic signature or profile of GIs tend to be highly correlated; and (ii) the corresponding genes are enriched for positive (suppressive) GIs¹. We observed similar trends in the mammalian data set using a set of known protein complexes and pathways (Supplementary Table 4). Indeed, pairs of genes whose products function together or are in the same complex were likely to interact positively (Fig. 3a) and have highly correlated phenotypic signatures (Fig. 3b). We also found that GI scores and profiles were strongly predictive of pairs of genes that are involved in the same functional module using an unbiased set of interactions based on probabilistic functional gene networks (HumanNet) at a range of cutoff values²² (Supplementary Fig. 8).

We observed enrichment for strongly negative GIs and PPIs between pairs of factors (Fig. 3a). These trends could be due to the fact that RNAi often only partially disables the function of a given protein. In these cases, an additional perturbation of another protein in the same module could result in complete abolishment of its function, and therefore a negative GI would be observed. Consistent with this reasoning is the observation that GIs in yeast derived from hypomorphic alleles of essential genes were more likely to produce negative GIs when the corresponding proteins are physically associated^3,11,23. Alternatively, these genetic relationships could correspond to non-essential genes whose proteins function in essential complexes. Additional data will need to be collected on a more global scale to study these relationships between PPI and GIs in more complex organisms. Finally, we observed a strong relationship between the fitness defect associated with a single gene knockdown and the number of GIs associated with that gene (Fig. 3c)^5,24,25, suggesting that genes with functional interactions with many other genes have the most dramatic effect on proliferative phenotypes. This is an important observation as it suggests that initial single RNAi screens could enable the design of more cost-effective, signal-rich E-MAPs.

A chromatin module map derived from the E-MAP

GI data from the E-MAP allow for the identification of individual functional modules at the level of individual S scores or by hierarchical clustering of GI profiles. For example, the five members of the PAF complex (encoded by Paf1, Leo1, Ctr9, Cdc73 and Rtf1), involved in transcriptional elongation and chromatin modification^26,27, display similar GI profiles and positive GIs with each other (Fig. 4a). In addition, Supt16h, a component of the FACT complex, a chromatin-specific elongation factor²⁸, was also in this functional module, consistent with the previous findings that the PAF complex is physically and functionally connected to FACT in budding yeast²⁷.

Our method also can be used to resolve functionally distinct modules in complexes. For example, members of the CNOT complex, involved in mRNA deadenylation and chromatin modification²⁹, form two distinct groups: a highly correlated, positively interacting module comprising products of Cnot1, Cnot2 and Cnot3 (CNOT core) and an uncorrelated group containing products of Cnot4, Cnot6, Cnot7 and Cnot8 (Fig. 4b). Consistent with this observation, a recent report demonstrated that Cnot1, Cnot2 and Cnot3 gene products behave as a functionally and physically distinct submodule within CNOT complex and the other components, including products of Cnot4, Cnot6 and Cnot8, bind to it in a dynamically regulated fashion³⁰. These results confirm that mammalian GI maps have the ability to dissect larger, functionally disparate complexes, as has been observed in simpler eukaryotic organisms².

Based on our manually curated set of protein complexes and pathways, we created a map depicting the genetic cross-talk between different functional modules represented in the E-MAP (Fig. 4c and Supplementary Tables 4 and 5). Contained within this map are modules corresponding to the INO80 chromatin remodeling complex³¹; the transcriptional regulator Integrator³²; the MCM DNA replication complex; and cohesin and condensin^33–35 complexes, involved in chromatin segregation and condensation. Notably, the PAF complex occupies a central position on the map, interacting strongly, both positively and negatively, with many other functional modules. For example, we observed strong positive GIs between the PAF complex and modules corresponding to CNOT and the RUVBL ATPase subcomplex (encoded by Ruvbl1, Ruvbl2 and Actl6a), which is part of several chromatin-remodeling complexes³⁶ (Fig. 4d). Furthermore, we observed negative interactions between components of the PAF complex and the PARP (encoded by Parp1 and Parp2) and condensin (encoded by Smc2 and Smc4) modules (Fig. 4d). To explore the genetic relationship between the PAF complex and PARP modules, we treated cells that are depleted for PAF complex function (using RNAi-mediated knockdown of Rtf1) with the PARP inhibitor, Veliparib (ABT-888)³⁷ (Supplementary Fig. 9). We observed a dose-dependent inhibition of growth specifically in the context of Rtf1 knockdown, consistent with the negative GIs we observed between PAF complex members and PARP (Supplementary Fig. 9). These data demonstrate that synthetic epistatic interactions can be leveraged to identify parallel pathways, which can be targeted using small molecules.

Additional investigation of the molecular basis of these individual connections will improve our understanding of how interactions among chromatin machines control the program of gene expression in mammals.

Validation of GIs using an orthogonal phenotypic readout

Phenotypic readouts that use growth rate to enumerate the proliferative capacity of a cell have been used in many GI screens^5,7,17. However, E-MAPs can potentially be constructed using any phenotype that can be observed by microscopy, including subtle effects on cell morphology or structure. Because the impact of one gene on another may differ depending on which pathways influence the phenotype being measured, E-MAPs generated using different readouts may not necessarily be similar to each other but are highly complementary in understanding gene function³⁸. However, we would expect partial overlap with our proliferation-based E-MAP and other phenotypic readouts because many different genes and pathways impinge on cells’ proliferation capacities. As the PAF complex is highly connected in the module map (Fig. 4c), we asked whether its epistatic interactions with other modules might be maintained when using alternate readouts. Using high-content microscopic interrogation, we observed that depletion of individual PAF-complex components leads to a very specific morphology phenotype in which cells display elongated nuclei (Fig. 5). This nuclear elongation may be due to disrupted regulation of the chromatin modification landscape as the PAF complex is required for normal histone H3 methylation at lysines 4, 36 and 79 (ref. 39).

We next tested the positive GIs that we had observed using the fitness-based readout between PAF complex components and the RUVB module. Combining a Ctr9 knockdown with knockdown of Ruvbl1, Ruvbl2 or Actl6a suppressed the elongated-nuclei phenotype (Fig. 5a,b), consistent with the positive interactions observed in our E-MAP (Fig. 4c,d). We did not see suppression of this specific phenotype when knockdown of Ctr9 was combined with knockdowns of other non-interacting factors, including Morf4l1 (Fig. 5 and data not shown), arguing there is a unique interaction between the PAF complex and the RUVBL module components. Although additional work will be required to understand why depletion of the PAF complex results in elongated nuclei and how absence of the RUVBL module suppresses this effect, these data demonstrate that we could recapitulate GIs with an orthogonal phenotypic readout.

Tissue culture

Standard media formulations were used throughout. Experiments were performed in 96- or 384-well tissue-culture plates (Greiner).

Cell lines

A mouse fibroblast cell line stably expressing nuclear-localized histone macroH2A-EGFP construct was used^41,42.

cDNA preparation

cDNA used for esiRNA preparation was prepared from mouse cells from RNA purified using Trizol (Invitrogen) and standard protocols. First-strand synthesis was primed with oligo(dT).

esiRNA preparation

esiRNA was prepared essentially as described previously¹⁵. Oligo sequences are listed in Supplementary Table 3. Below is a detailed description as this is a multistep process and some of the steps were optimized for 96-well format preparations.

Array preparation

We dispensed 7.5 ng of an array of 96 esiRNAs at 0.75 ng/μl in 384-well tissue-culture plates in quadruplicate using a Biomek FX liquid handler. Then, 7.5 ng of the query esiRNA was dispensed into each well using a Multidrop Nano (Thermo) liquid dispenser attached to a custom-built liquid handler (Supplementary Fig. 1). The plates were vacuum-dried and could be stored without apparent loss of activity until transfected.

Transfection of X3 macroH2A-EGFP cells

Lipid-based transfection using Lipofectamine RNAi MAX (Invitrogen) was used. For 384-well format tissue-culture plates 750–1,000 cells/well were transfected with 15 ng (7.5 ng each) esiRNA using 0.1 μl transfection reagent per well. Lipid-RNA complexes were reconstituted in OptiMEM for 15–30 min before cell addition. Cells were added in antibiotic-free DMEM supplemented with 10% FBS and incubated overnight. Before we sealed the plates, fresh DMEM supplemented with penicillin, streptomycin and 10% FBS was added.

Quantitative reverse transcription–PCR

cDNA was prepared using CellSure cDNA kit (Bioline) per manufacturer’s instructions. Quantitative reverse transcription (RT)-PCR was carried out using SensiFAST SYBR lo-ROX master mix (Bioline) on a MX4000 qPCR instrument (Stratagene).

Raw-data collection

Cell counting was done using an Acumen eX3 plate cytometer (TTP Labtech) using 384-well tissue culture plates. Data were collected in batches of more than 20 screens each.

Scoring of GIs

Raw data were scored using a published software toolbox²⁰. Individual batches were normalized and scored separately, thus minimizing systematic experimental biases and batch-to-batch variation.

Detailed explanation of the S score

Data processing using E-MAP toolbox

To launch the graphical user interface, “emapGUI” was typed at the command prompt in Matlab. The software offers several different preprocessing and normalization options. We used the options menu to set those. The following were used for processing the raw data: normalization method, Parzen; control size method: Parzen; preprocessing option: no preprocessing; threshold for removing frequently missing data: 0.15; filter data with missing neighbors: no filter. We used the ‘File’ --> ‘load raw data’ command to load data from source files (see the included example files for templates for making the necessary coordinate and file name maps). We used ‘Data processing … score data’ to generate unaveraged S scores. We used ‘File’ --> ‘Save as MAT file’ to save the analysis. Individual batches were scored separately, and the resulting score tables (scoremat structure in the .mat file) were then merged to create the final score matrix. Data for replicate screens that correlated at ≤ 0.1 were removed from the dataset.

HCS microscopy

A Cell Insight Personal Image Cytometer (Thermo) with a 10× objective was used to collect microscopy data. Image analysis was performed using the Cellomics vHCS Discovery Toolbox (Thermo). Nuclear shape in Figure 5 was defined as a length to width ratio for each cell.

Module map

Curated protein complexes and pathways were integrated withS scores from this study. Between-module interactions were assessed by comparison of the average S score of GIs between two modules to 10,000 sets of random interactions of the same size as in ref. 23. Displayed edges had P < 0.01.