Articles
nAture methods | VOL.8 NO.4 | APRIL 2011 | 341
the analysis of synthetic genetic interaction networks can
reveal how biological systems achieve a high level of complexity
with a limited repertoire of components. studies in yeast and
bacteria have taken advantage of collections of deletion strains
to construct matrices of quantitative interaction profiles and
infer gene function. Yet comparable approaches in higher
organisms have been difficult to implement in a robust manner.
here we report a method to identify genetic interactions
in tissue culture cells through rnAi. By performing more
than 70,000 pairwise perturbations of signaling factors,
we identified >600 interactions affecting different quantitative
phenotypes of Drosophila melanogaster cells. computational
analysis of this interaction matrix allowed us to reconstruct
signaling pathways and identify a conserved regulator of
ras-mAPK signaling. large-scale genetic interaction mapping
by rnAi is a versatile, scalable approach for revealing gene
function and the connectivity of cellular networks.
Genetics underlying many phenotypes, including most common
diseases, are complex with contributions from multiple loci. Studies
in model organisms provide evidence for pervasive genetic inter
actions with large effects on many phenotypes. Such genetic factors
are difficult to identify in classical lossoffunction screens. For
example, RNAi screens in cultured cells have revealed components
of signaling pathways through pathwayspecific reporter assays but
revealed little about the interactions between the different com
ponents. To explore the underlying network connectivity, simul
taneous perturbations of multiple components are required, for
example, through combinatorial drug treatments1 or the genera
tion of double mutant strains2,3. Here we describe an RNAibased
experimental approach to functionally annotate metazoan genes
based on their genetic interaction profiles, independent of path
wayspecific reporters, mutant collections or chemical inhibitors.
results
double-rnAi analyses
RNAi offers the opportunity to simultaneously reduce the expres
sion of any chosen pair of genes, allowing us to systematically
mapping of signaling networks through synthetic
genetic interaction analysis by rnAi
Thomas Horn1,2,5, Thomas Sandmann1,3,5, Bernd Fischer4,5, Elin Axelsson4, Wolfgang Huber4 & Michael Boutros1
sample large numbers of distinct, biologically relevant conditions.
To overcome limitations of RNAi experiments, such as offtarget
effects, we designed two independent dsRNAs to each target
and used robust statistical modeling to identify singleRNAi or
doubleRNAi phenotypes and synthetic effects. As genetic inter
actions can manifest themselves by affecting any phenotypic trait,
we used automated microscopy to collect multiparametric data,
for example, information about cell number, nuclear size and
fluorescence intensity after Hoechst staining, instead of restricting
the analysis to any individual, preselected pathway4. To validate
our approach, we assessed genetic interactions between known
components of the RasMAPK pathway affecting cell number at
varying concentrations of dsRNA (Fig. 1a–c and Supplementary
Fig. 1). Depletion of the Drosophila Ras85D gene, which encodes
a member of the conserved Ras superfamily of small GTPases, led
to a concentrationdependent reduction in cell growth5,6 (Fig. 1a).
RNAimediated knockdown of CG13197, a gene not connected
to Ras signaling, also attenuated cell growth. When we simulta
neously targeted both Ras85D and CG13197 with different con
centrations of dsRNAs, we could predict the resulting number of
cells based on the singleknockdown phenotypes of the individual
genes by a multiplicative model (Fig. 1a,d), as has been observed
for double deletion strains in yeast2,7. Deviations between the
expected and the experimentally observed doubleRNAi effects
revealed positive (alleviating) or negative (aggravating for genes
with negative singleRNAi effect) genetic interactions (Fig. 1d).
The outcome of simultaneously targeting two genes with known
functions in the RasMAPK signaling pathway deviated strongly
from predictions based on singleknockdown phenotypes. Double
RNAi of Ras85D and Gap1, a negative regulator of Ras, alleviated
the growth inhibition caused by targeting Ras85D alone, revealing
a strong positive interaction (Fig. 1b). Double RNAi of Gap1 and
Ptp69D, a receptorlinked protein tyrosine phosphatase, in con
trast, led to a negative genetic interaction (Fig. 1c). Thus, double
RNAi can reveal both positive and negative interactions, whose
strengths, analogous to interactions between chemical com
pounds8, are concentrationdependent and reflect the quantitative
nature of functional interactions (Supplementary Fig. 1).
1German Cancer Research Center (Deutsches Krebsforschungszentrum), Division Signaling and Functional Genomics and Heidelberg University, Department of Cell
and Molecular Biology, Faculty of Medicine Mannheim, Heidelberg, Germany. 2Heidelberg University, Hartmut HoffmannBerling International Graduate School
for Molecular and Cellular Biology, Heidelberg, Germany. 3Heidelberg University, CellNetworks Cluster of Excellence, Heidelberg, Germany. 4European Molecular
Biology Laboratory, Genome Biology Unit, Heidelberg, Germany. 5These authors contributed equally to this work. Correspondence should be addressed to
M.B. (m.boutros@dkfz.de) or W.H. (whuber@embl.de).
Received 23 SeptembeR 2010; accepted 4 FebRuaRy 2011; publiShed online 6 maRch 2011; doi:10.1038/nmeth.1581
©
2
01
1
N
at
u
re
A
m
er
ic
a,
In
c.
A
ll
ri
g
h
ts
r
es
er
ve
d
.
342 | VOL.8 NO.4 | APRIL 2011 | nAture methods
Articles
high-throughput analysis of pairwise interactions
We tested all pairwise interactions between 93 genes involved
in signal transduction in Drosophila cells, evaluating two non
overlapping RNAi reagents for each target (192 dsRNA reagents
including controls; Fig. 1e, Supplementary Figs. 2 and 3, and
Supplementary Table 1)9. Targeted genes included annotated
components of the three MAPK pathways (RasMAPK, JNK and
p38 pathway) and all annotated protein and lipid phosphatases
expressed in Drosophila Schneider cells (Supplementary Table 2).
We performed experiments in Schneider S2 cells, which we fixed,
stained with Hoechst dye and analyzed using highthroughput
fluorescence imaging and automated image analysis (Fig. 1e).
We selected three nonredundant quantitative features from the
images: number of cells per well, mean nuclear area and nuclear
fluorescence intensity (Fig. 1e and Supplementary Fig. 4). For
example, treatment with dsRNA to the firefly luciferase gene
(negative control) led to an average of 48,200 cells per well with a
mean nuclear area of 59.1 µm2 per cell. Depletion of Rho1, a small
GTPase involved in cytokinesis and cytoskeleton remodeling10,
led to significantly larger nuclear area (77.9 µm2 per cell, P < 1 ×
10−15, Student’s ttest, n = 16), likely reflecting multinucleate
morphology caused by incomplete cytokinesis, with a concomi
tant decrease in the number of cells (25,400 cells per well, P < 1 ×
10−15, Student’s ttest, n = 16). In contrast, pnt depletion also
resulted in a decreased number of cells (19,000 cells per well, P < 1 ×
10−15, Student’s ttest, n = 16) but a smaller mean nuclear area
(44.7 µm2 per cell, P < 1 × 10−15, Student’s ttest, n = 16).
We performed two biologically independent experiments
yielding 73,728 measurements in total, from which we esti
mated interaction scores (Fig. 2a, Supplementary Figs. 5–8 and
Supplementary Table 3). Phenotypic measurements were highly
a b c
2–2
π score
e
d FlucA B
100%80% 62.5%
75% 100%50%12.5%
–2 +10
Aggravating
(negative)
Alleviating
(positive)
P
he
no
ty
pe
In
te
ra
ct
io
ns
25%
Interaction
score (πA,B)
CG13197 dsRNA (ng)
R
as
85
D
d
sR
N
A
(
ng
)
0 50 100
0
50
100
0.0
Gap1 dsRNA (ng)
R
as
85
D
d
sR
N
A
(
ng
)
0 50 100
0
50
100
0.5
1.0
1.5
2.0
Gap1 dsRNA (ng)
P
tp
69
D
d
sR
N
A
(
ng
)
0 50 100
0
50
100 −1.
5
−1.0
−0.5
0.0
...
N
um
be
r
A
re
a
In
te
ns
ity
Interaction
score (πA,B)
+ RNAi 1
+ RNAi 2
+ RNAi 192
Combinatorial RNAi
Imaging and
image analysis
Modeling of
genetic interactions
Gene A Gene B
A1
A2
B1
B2
×
Figure 1 | A multiparametric approach
to identify genetic interactions through
double-RNAi. (a–c) Genetic interaction
surfaces of double-RNAi treatments over
a range of dsRNA concentrations. Axes
indicate the amounts of the respective
dsRNAs combined per well. Interaction scores (π scores
based on cell number phenotypes) are shown on a color
scale ranging from −2 (negative interaction) to 2 (positive
interaction). (d) Schematic overview of π score calculation.
Single RNAi effects (A and B) are compared to that of a negative control dsRNA targeting firefly luciferase (Fluc). The expected double-RNAi effect is
obtained by multiplying the single RNAi effects (arrowhead points to the relative cell number of 50% expected in this example) and compared to the
observed double-RNAi phenotype. The π score is the log2 ratio between the observed and the expected value. (e) Schematic overview of the combinatorial
RNAi experiment. Each color corresponds to a single dsRNA in the assay plates. To each plate, a different second dsRNA (RNAi 1–192) is added to all wells.
This design creates all possible dsRNA combinations (arrows) targeting each pair of genes, A and B, with two dsRNAs (A1 and A2, and B1 and B2).
c
a
b
rl
phl
csw
msk
Sos
Ras85D
Dsor1
Cka
stg
mRNA−cap
mop
PpV
CG3573
Gap1
Rho1
pnt
puc
drk
Pvr
mts
π−
sc
or
e
nu
cl
ea
r
ar
ea
π−
sc
or
e
ce
ll
nu
m
be
r
Jra
kay
shark
slpr
bsk
π−
sc
or
e
m
ea
n
in
te
ns
ty
−0.5
0.5
−0.2
0.2
−0.2
0.2
Figure 2 | Clustering of genetic interaction profiles predicts gene
function. (a) Hierarchical clustering of the genetic interaction profiles
based on observed cell number. Known signaling components from
the Ras-MAPK pathway (top right) and the JNK pathway (bottom right)
are highlighted. mRNA-cap, gene encoding mRNA-capping enzyme.
(b,c) Genetic interaction profiles for Ras-MAPK (right; dark gray) and
JNK (right; light gray) regulators based on nuclear area per cell (b) or
mean signal intensity (c). Genes in b and c are ordered as in a.
©
2
01
1
N
at
u
re
A
m
er
ic
a,
In
c.
A
ll
ri
g
h
ts
r
es
er
ve
d
.
nAture methods | VOL.8 NO.4 | APRIL 2011 | 343
Articles
reproducible across technical and biological replicates, with a
Pearson correlation coefficient of 0.95 for observed numbers
of cells between biological replicates (Supplementary Fig. 9).
In each replicate, we measured the doubleRNAi phenotype for
each pair of genes eight times (Fig. 1e), allowing for stringent
statistical analysis.
Interactions based on cell number, nuclear area or intensity
were clearly distinct (Fig. 2a–c and Supplementary Figs. 6 and 7).
For example, depletion of downstream of receptor kinase (drk), a
positively acting component of the RasMAPK signaling pathway,
which provides mitogenic signals in S2 cells6, caused reduced cell
growth as well as a reduction in cell size (Fig. 3 and Supplementary
Table 3). Rho1 RNAi knockdown, in contrast, caused cyto
kinesis defects, resulting in large, multinucleated cells (Fig. 3).
The observed singleRNAi phenotypes of drk and Rho1 there
fore differed depending on the phenotype studied: both led to a
reduction in cell numbers but caused opposite effects on cell size.
As the cytokinesis defects caused by Rho1 RNAi knockdown can
not manifest themselves in the absence of mitogenic signals from
the RasMAPK pathway, the drk-Rho1 doubleRNAi phenotype
(Fig. 3) was similar to the drk singleRNAi situation: Rho1 is epi
static to drk. Comparing the observed quantitative doubleRNAi
phenotypes with those expected under a multiplicative model
revealed deviations for both phenotypes: although the cells were
smaller than expected (resulting in a negative interaction), there
were more cells than predicted (a positive interaction). Thus,
multiparametric phenotyping revealed contextspecific inter
actions, affecting different phenotypes with different strength
and/or direction.
Interaction scores were quantitatively reproducible with a
Pearson correlation coefficient of 0.62 between biological repli
cates (Supplementary Fig. 10). We accepted interactions with
local false discovery rate (FDR) of 5% for each phenotype and
identified 634 interactions in total. Of these, 372 interactions
affected the number of cells, 379 led to changes in nuclear area
per cell and 337 modulated the nuclear fluorescence intensity
(Fig. 4a and Supplementary Table 3). We validated a subset of
interactions that affected the number of detected nuclei using an
independent, enzymatic assay for cell viability. The interactions
were highly reproducible irrespective of the method used to
assess the number of cells per well (Supplementary Fig. 11).
Although we observed 135 interaction pairs for all phenotypes,
315 (49.7%) were specific to a single phenotypic readout, high
lighting the multidimensionality of the genetic interaction space.
For validation, we compared the identified genetic interactions
with previously reported interactions in Drosophila or between
human interologs11. Both the common as well as the pheno
typespecific subsets of genetic interactions were significantly
enriched for annotated interaction pairs (Fisher’s exact test,
Pvalues between 10−14 and 0.05; details and P values are available
in Supplementary Fig. 12). For example, we observed a genetic
interaction affecting cell number between drk and the inositol
5′phosphatase synaptojanin (synj), whose orthologs GRB2 and
SYNJ1/2 associate with each other in human cells12 In addi
tion, components of the JNK and Ras/MAPK pathways known
to associate physically were significantly more likely to interact
genetically (P < 0.01, Fisher’s exact test; Supplementary Fig. 13).
The numbers of detected positive and negative interactions were
approximately the same (Fig. 4b and Supplementary Fig. 14). In
particular, we observed a high frequency of interactions between
known components of the RasMAPK signaling pathways, reflect
ing the focused design of this dataset.
Only 15 genes lacked any interactions affecting the number of
detected cells at an experimentwide FDR of 5% (Fig. 4b). For
example, RNAi knockdown of the protein phosphatase CG10376
by itself caused a reduction in cell number (Fig. 4c). The observed
pairwise phenotypes of CG10376 with all other genes, including
those with strong positive or negative singleRNAi phenotypes,
were consistent with predictions from combining single gene
effects and did not indicate the presence of a genetic interac
tion (Fig. 4c). In contrast, depletion of Gap1 alleviated the RNAi
phenotypes of many other components of the RasMAPK path
way (Fig. 4d), often restoring approximately normal growth.
Genetic interaction profiles
In addition to revealing individual pairwise genetic interactions,
we obtained an informative genetic interaction profile, a vector
of interactions with all targeted loci, for each gene. Unsupervised
clustering of these profiles reconstructed known global and local
relationships between the assayed genes without requiring any
Figure 3 | Phenotype-specific genetic interactions. (a,b) Single-RNAi
and double-RNAi effects of targeting Rho1 and/or drk on nuclear
area (a) or cell number (b). Observed and expected phenotypes are
indicated as percentage relative to the negative control treatment (Fluc
dsRNA). (c) Schematic representation of drk and/or Rho1 RNAi effects on
nuclear area and cell number. Images are fluorescence microscopic images
of S2 cells after RNAi treatment, stained with Hoechst and antibodies to
α-tubulin. Scale bars, 10 µm.
c
drk Rho1
drk Rho1
drk Rho1
drk Rho1
DNA
α-tubulin
a b
100
Nuclear area
(percentage of negative control)
500
Cell number
(percentage of negative control)
150
100%
81%
131%
106% (expected)
83% (measured)
100%
34%
53%
18% (expected)
34% (measured)
50 1000
Fluc
drk
Rho1
drk and Rho1
drk and Rho1
©
2
01
1
N
at
u
re
A
m
er
ic
a,
In
c.
A
ll
ri
g
h
ts
r
es
er
ve
d
.
344 | VOL.8 NO.4 | APRIL 2011 | nAture methods
Articles
additional information (Fig. 2a and Supplementary Figs. 5–7).
We observed clear separation of components of the RasMAPK
pathway from other signaling components, such as regulators of
JNK signaling (Fig. 2a). Notably, the JNK phosphatase puckered
(puc) was grouped together with positive regulators of the Ras
MAPK pathway. In addition to its role as a negative feedback regu
lator of the JNK pathway, puc has also been shown to be required
for maximum activity of Rolled (rl) in Drosophila tissue culture
cells13, and this signaling crosstalk is reflected in our dataset.
cka is a modulator of ras-mAPK signaling
To assign genes to known signaling pathways, we trained a clas
sifier on the combined, multiparametric interaction profiles of
annotated components of the RasMAPK and JNK pathways and
estimated its predictive power through crossvalidation (Fig. 4e).
The classifier correctly identified most of the known positive
regulators of RasMAPK signaling in the dataset (for example,
the PP2A catalytic subunit microtubule star (mts) or the tyrosine
phosphatase myopic (mop)) (Fig. 4e,f) and clearly separated them
from negative regulators. Classifications of the latter overlapped
partially with components involved in JNK signaling, reflecting
the antagonistic relationship between the two pathways13. In addi
tion, we observed several unexpected functional relationships.
For instance, connector of kinase to AP1 (encoded by Cka),
previously described as a scaffold protein in the JNK signaling
pathway14, was predicted with high confidence to act as a positive
regulator of RasMAPK signaling. We therefore characterized and
validated the role of Cka in this pathway through independent
biochemical and genetic experiments. Similar to known regu
lators of RasMAPK signaling, Cka genetically interacted with
individual components of both the RasMAPK and JNK path
ways, and its correlation profile was highly similar to that of
Ras85D (Fig. 5a and Supplementary Fig. 15). As knockdown of
Ras85D or rl, depletion of Cka led to a significant reduction in
mRNA levels of sprouty (sty)15 (Fig. 5b; P < 0.05, Student’s ttest,
n = 6) and affected other genes downstream of the RasMAPK
cascade (Supplementary Fig. 16). In contrast, neither RNAi
knockdown of bsk nor slpr, the Drosophila JNK and JNKKK,
caused any substantial change in sprouty (sty) expression.
Next, we investigated whether RNAi to Cka affected the activ
ity of the RasMAPK cascade. Knockdown of Cka led to reduced
basal RolledERK phosphorylation levels in Drosophila S2 cells
(Fig. 5c and Supplementary Fig. 17). A similar attenuation
of RasMAPK pathway activity was observed in human cells
upon RNAi knockdown of Cka’s orthologs, Striatin (STRN) or
Striatin3 (STRN3) (Fig. 5d and Supplementary Figs. 18 and 19).
Drosophila GCKIII, a modifier of Ras/MAPK signaling13, as well
as its binding partner, the PP2A catalytic subunit Microtubule
star (mts), immunoprecipitated along with Cka (Fig. 5e and
Supplementary Figs. 20 and 21). Cka also has a role downstream
of the epidermal growth factor receptor (Egfr) in vivo, as reducing
Cka gene dosage in the background of the gainoffunction allele
EgfrElpB1 partially suppressed the formation of ectopic wing vein
material (P < 1 × 10−10, Fisher’s exact test; Fig. 5f). Drosop