ArticlePDF AvailableLiterature Review

Single-cell and multivariate approaches in genetic perturbation screens



Large-scale genetic perturbation screens are a classical approach in biology and have been crucial for many discoveries. New technologies can now provide unbiased quantification of multiple molecular and phenotypic changes across tens of thousands of individual cells from large numbers of perturbed cell populations simultaneously. In this Review, we describe how these developments have enabled the discovery of new principles of intracellular and intercellular organization, novel interpretations of genetic perturbation effects and the inference of novel functional genetic interactions. These advances now allow more accurate and comprehensive analyses of gene function in cells using genetic perturbation screens.
Large-scale genetic perturbation screens have been
instrumental in many biological discoveries1–3. These
screens use perturbations that act at the DNA, RNA
(post-transcriptional) or protein (post-translational)
level, providing a variety of different readouts. Given
their fundamental importance in genetics, molecu-
lar cell biology and systems biology, these methods
— as well as the various commonly applied statistical
approaches to extract information from large-scale
genetic perturbation screens4–8 — have been extensively
described previously1–3,6,9,10.
In this Review, we do not elaborate on methods for
genetic screens, although we provide an overview of
the relevant techniques. We specifically focus on two
related aspects, the importance of which only became
apparent recently. First, we describe the phenomenon
of cell‑to‑cell variability or cellular heterogeneity (BOX1),
which is a fundamental property of populations of cells,
and discuss recent advances in the ability to quantify,
at a large scale, multiple parameters of genetic pertur-
bation effects in thousands of single cells1116. Second,
we describe the general concept of using quantitative
multivariate readouts from large-scale genetic pertur-
bation screens to infer functional interactions between
phenotypic properties and between genes. Finally, we
present an outlook on some of the future opportunities
that the single-cell paradigm will bring to the unravel-
ling of biological complexity from large-scale genetic
perturbation screens.
Genetic perturbation screens
Traditionally, genetic perturbation approaches relied
on random perturbations of the DNA of an organism
or cells using chemical mutagens or random inser-
tions2,17, and are also termed forward genetics. These
approaches create a null or mutated allele, with the
latter causing either a constitutive or, sometimes, a
conditional mutation such as a temperature-sensitive
mutant protein. In the past decade, sequence-specific
genetic perturbations (also termed reverse genetics),
such as post-transcriptional gene perturbations by
means of RNA interference (RNAi), have increasingly
been performed, allowing large-scale targeted knock-
down of specific mRNAs in Caenorhabditiselegans
and Drosophilamelanogaster, as well as in mammalian
cells10,18–27. In addition, over expression screens of either
wild-type or mutated forms of genes, usually encoded
as cDNAs from plasmids, have also been applied28,29.
Chemical compound or inhibitor screens (also termed
chemical genetics’)30, which usually rely on post-
translational perturbations in which the activity or func-
tion of the protein is inhibited by a small molecule,are
also now frequently applied11,30. Furthermore, there
aremultiple genome-editing approaches that target spe-
cific regions of the genome to create a null or mutated
allele31–33. These genome-editing approaches have now
become more efficient and can be applied at a large
scale34–39. Recently, specific genome targeting approaches
using single guide RNAs (sgRNAs) for the CRISPR–Cas9
1Institute of Molecular Life
Sciences, University of Zürich,
Winterthurerstrasse 190,
CH-8057 Zürich, Switzerland.
2Present address: CeMM
Research Center for
Molecular Medicine of the
Austrian Academy of
Sciences, A-1090 Vienna,
Correspondence to L.P.
e-mail: lucas.pelkmans@imls.
Published online
2 December 2014
Cell‑to‑cell variability
The phenomenon that
individual cells in a population
of genetically identical cells
display variable activities and
Single‑cell and multivariate
approaches in genetic perturbation
Prisca Liberali1, Berend Snijder1,2 and Lucas Pelkmans1
Abstract | Large-scale genetic perturbation screens are a classical approach in biology
and have been crucial for many discoveries. New technologies can now provide
unbiased quantification of multiple molecular and phenotypic changes across tens
of thousands of individual cells from large numbers of perturbed cell populations
simultaneously. In this Review, we describe how these developments have enabled
the discovery of new principles of intracellular and intercellular organization, novel
interpretations of genetic perturbation effects and the inference of novel functional
genetic interactions. These advances now allow more accurate and comprehensive
analyses of gene function in cells using genetic perturbation screens.
Nature Reviews Genetics
AOP, published online 2 December 2014; doi:10.1038/nrg3768 REVIEWS
© 2014 Macmillan Publishers Limited. All rights reserved
Cellular heterogeneity
Similar to cell‑to‑cell variability.
Sometimes, ‘heterogeneity’
is used to indicate multiple
discrete phenotypes within a
population, while ‘variability’
is used to indicate variation
around a single phenotype.
There is no consensus on which
term to use in which occasion,
and both terms are
Multivariate readouts
Phenotypic readouts consisting
of multiple features of the
cellular activity, state and
Functional interactions
A general term that
incorporates protein–protein
interactions, classical
genetic interactions, regulatory
interactions (such as kinase–
substrate interactions) and
phenotypic interactions.
(clustered regularly interspaced short palindromic
repeat–CRISPR-associated protein9) system have been
repurposed to induce sequence-specific repression or
activation of gene expression at a genome scale39,40.
Different approaches have different effects. Every large-
scale gene perturbation approach has its advantages and
disadvantages (TABLE1). In general, with most approaches
there is a trade-off between specificity and duration in
establishing a measurable perturbation41,42. Importantly,
differences in the method of perturbation and the time
required to establish the perturbation can result in
differences in the observable effects. Another factor to
bear in mind is that some enzymes can sufficiently per-
form their task in a cell even when they are reduced to a
fraction of their normal concentration and, in this case,
knockdown by RNAi may not result in a measurable
effect, whereas a gene deletion will do so. In addition,
the presence of an inhibited protein in a cell can have
a different effect from that of the absence of the same
protein42. Changes observed in a cell population after
a few days of gene knockdown can also be very differ-
ent from those observed in a cell population selected
over the course of several weeks to harbour a specific
Box 1 | Cell-to-cell variability and the microenvironment
Cell‑to‑cell variability refers to the phenomenon that no two genetically identical cells have identical behaviour and
appearance. The extent and origins of cell‑to‑cell variability depend on the cellular activity that is compared between
cells, but its study has revealed several common trends88,95,141–144. Chance, or stochasticity, has a considerable role in
cellular processes that involve a small number of molecules. Small differences can lead to sizeable differences in cellular
behaviour. However, in general, robustness in molecular mechanisms145–149 buffers the intrinsic stochasticity of molecular
processes, and extrinsic factors are found to explain the majority of total cell‑to‑cell variability. One major component of
such extrinsic factors, particularly in adherent cells, is the microenvironment of individual cells. Even in environmentally
controlled cell culture conditions, a growing population of adherent cells will be continuously subjected to changing
microenvironments as a consequence of an increase in cell number together with increased cell adhesion and
migration79. As the population size increases, so does the local cell density but with different rates for each single cell,
and more cells will find themselves entirely surrounded by neighbouring cells. Genetic perturbations can alter the
distribution of microenvironmental properties of single cells (see the figure; Perturbations A and B) to such an extent
that they dominate the effect of a perturbation in genetic perturbation screens14,16. Although it is common practice to
normalize cellular readouts obtained in high‑throughput approaches for differences in the total cell number, the
population context of individual cells can be substantially different even for populations with equal number of cells
(see the figure). Therefore, multiparametric methods are required that correct for the influence of the population
context at the single‑cell level14. Owing to the technical challenges associated with measuring single‑cell behaviour
quantitatively for whole cell populations, many questions remain. There is no comprehensive understanding of which
cellular processes are influenced by the population context, and few links between population context‑dependent cellular
processes and the invivo multicellular programming of cells have been investigated so far150,151.
Total cell number
Number of cells facing
empty space
Number of cells
Edge cells (%)
Nature Reviews | Genetics
Relative abundance
Relative abundance
0 5 10 15 0 5 10
Nucleus diameter (µm) Local cell density
Perturbation BPerturbation A
© 2014 Macmillan Publishers Limited. All rights reserved
Synthetic interaction screens
Genetic screens in which two
perturbations are combined to
assess the possible synergistic
and epistatic effects between
the two genes perturbed.
Mammalian haploid cells
Mammalian cells that harbour
only one copy of the genome.
null allele and hence an adaptive phenotype. All of these
factors need to be considered when designing genetic
perturbation screens and analysing data produced from
differing approaches.
Arrayed versus pooled screens. An important difference
between genetic perturbation screens is whether the per-
turbation is applied all at once (pooled) or one by one
(arrayed)3,4,10,15,26,27,43–46 (FIG.1a). In pooled screens, a mix-
ture of single genetic perturbations — such as barcoded
short hairpin RNAs (shRNAs)47,48, sgRNAs for Cas9
(REFS36,37) or random insertion of gene-trap cassettes49
— or double genetic perturbations (using, for example,
vectors that express 2 shRNAs48) is applied to one large
population of cells. After positive selection of cell clones
with a desired trait — such as resistance to a cytotoxic
compound or pathogen, or presence of a selectable
reporter — the original perturbations can be retrieved by
means of sequencing, which allows the identification of
the exact location of the genetic perturbation in each
of the selected clones. Pooled screens are easy to carry
out and therefore allow substantial up-scaling of the
number of genetic perturbations tested, including
the feasibility to carry out synthetic interaction screens at a
large scale48. However, pooled screens rely on the use of
a selective pressure, or the sorting of cells with a desired
signal, to select relevant cells from a complex pool of
perturbed cells, and on subsequent enmasse identifi-
cation of all selected perturbations (FIG.1a). Therefore,
perturbations that affect cell proliferation or viability
are usually lost, as are perturbations that do not confer
complete resistance to a selective pressure. In addition,
pooled approaches currently do not have single-cell
resolution and cannot obtain multivariate information
from thousands of single cells subjected to the same per-
turbation, and they are not, or only to a limited extent,
quantitative (FIG.1b). By contrast, genetic screens in
arrayed format do not have these disadvantages and can
have numerous quantitative multivariate dimensions in
the identification of hits and in the inference of genetic
interaction (see Supplementary information S1 (table)
for an overview of the multivariate dimensions currently
used in arrayed screens and methods used for reducing
dimensionality in single-cell experiments).
Large-scale collections of single-gene deletions. For sev-
eral single-cell organisms, genome-wide collections of
strains with viable single-gene deletions are available50–52.
For mammalian organisms, such genome-wide resources
do not currently exist, but there are efforts to create
genome-wide collections of viable single-gene knock-
out mice and their cell lines53–55. Technologies of higher
throughput are also emerging, which rely on the use
of random mutational insertions in mammalian haploid
cells49,56–58 or gene-editing methods in mammalian
haploid or diploid cells34,36,37,39. CRISPR–Cas9-mediated
gene editing shows high efficiency in diploid cells and
is easily targeted to a specific site in the genome by an
sgRNA, and it is likely that this technology will become
the method of choice for large-scale gene knockout
screening in a variety of mammalian cells36,37,39, includ-
ing tissue culture cell lines, primary somatic cells and
stem cells. In creating such large single-gene deletion
collections, the biggest challenge will be to grow and
assay each cell line in such a collection in parallel,
which will require substantial efforts and more sophis-
ticated automation and liquid-handling robotics than
those currently used by most academic laboratories.
Eventually, large-scale genetic perturbation screens
in human cells may become most powerful when
they can be applied in an arrayed format and rely on
null deletions. A rapid, high-throughput one-by-one
gene deletion approach in multiwell plates without the
need for selection, similar to small interfering RNA
(siRNA) and small compound screen methods, could
provide such a solution in thefuture.
Table 1 | Advantages and disadvantages of different genetic perturbation methods
Perturbation Level of
Advantages Disadvantages
Haploid screens DNA (random
High specificity Depends on selective pressure or sorting;
cannot achieve single-cell resolution
DNA High specificity Inefficiency in production; there can be
adaptive mechanisms and off-target effects
DNA Experimental inference of
genetic interactions
Exponential increase in experiment size
gene knockout
DNA High specificity and high
efficiency of gene knockout
Adaptive mechanism and possible
off-target effects
High specificity and few
off-target effects
RNAi screens (siRNA
and shRNA screens)
mRNA Experimentally accessible
way to perform arrayed
Off-target effects and incomplete
knockdown efficiency
Compound screens Proteins Short time of action Poor specificity; not genome-wide
CRISPR, clustered regularly interspaced short palindromic repeat; RNAi, RNA interference; shRNA, short hairpin RNA; siRNA, small
interfering RNA.
© 2014 Macmillan Publishers Limited. All rights reserved
Double-gene perturbations. In principle, any combina-
tion of gene perturbation methods can be applied to
study how the combined effect of two perturbations
aggravates or alleviates their respective single pertur-
bation effects (FIG.1c). Testing these epistatic effects
between two gene perturbations is a classical approach
in genetics, as it can reveal genes co-functioning in the
same pathway or protein complex, but its large-scale
application has traditionally been reserved to yeast59–61.
However, approaches combining RNAi and/ or chemi-
cal compound screening to evoke double-gene per-
turbations in mammalian cells48,62–65, as well as the
combination of large-scale insertional mutagenesis
screening in mammalian haploid cells with cytotoxic
drugs66, are now increasingly being used. Clearly, pair-
wise double-gene perturbation screens rapidly become
an enormous challenge; for example, 400 single-gene
perturbations have 79,800 possible pairwise combina-
tions. Double-gene perturbation is therefore usually
either restricted to a preselected set of genes for which it
would be interesting to comprehensively map their pair-
wise epistatic effects or applied in a pooled screening
format48,67,68 (see above).
Interpretation of genetic perturbation effects
Obviously, the interpretation of a genetic perturba-
tion effect relies primarily on the readout. Given the
complexity of cellular activities, the number of genes
involved and the cell-to-cell variability that these activi-
ties can display, it is now becoming clear that measuring
multiple aspects of cellular phenotype in many individ-
ual cells is necessary to avoid sampling bias or incor-
rect interpretation of perturbation effects. In addition,
single-cell distributions of multivariate readouts allow
sensitive detection of perturbation effects that occur in
only a subset of cells and offer a better characterization
of the effectitself.
Obtaining multivariate single-cell readouts. There are
numerous ways to obtain a multivariate set of measure-
ments from a large number of single cells, but two main
approaches are practical when applied in large-scale
genetic perturbation screens (FIG.1b). These are flow
cytometry and high-throughput imaging, and both
methods have undergone rapid development.
In flow cytometry, the use of antibodies labelled
with heavy isotope combinations and mass spectros-
copy as the method of detection has allowed a marked
increase in the extent of multiplexing69. This approach
is termed mass cytometry (also known as CyTOF)69 and
has, in various studies, allowed the multiplexing of up
to 35 different molecular readouts from thousands of
single cells70,71. So far, this approach has been primar-
ily applied to quantify the levels of signalling proteins
and their phospho-specific modifications in single cells
to reveal molecular heterogeneity in cancer, as well as
hetero geneous adaptation of signalling during immune
and drug responses70,71.
In imaging, there have been important developments
in computational methods to segment single cells within
images of cells grown in culture or within their context
1° DNA-level
(e.g. gene knockouts)
Primary perturbation
dimension (1°) Pooled versus arrayed screening Readouts
t > several days
t < hours or days
1° DNA/RNA-level
(e.g. siRNA)
1° Protein-level
(e.g. drugs)
1° DNA/RNA-level
(e.g. siRNA)
1° RNA-level
(e.g. siRNA)
High-content dimension (3°)
Single-cell phenotypes
Single-cell transcriptomics
(Drugs, compounds
and peptides)
(siRNA, shRNA
and miRNA)
Pooled cells
Throughput (conditions measured/time)
Plate reader
Automated microscopy
(Gene trap and
Plate reader
FACS and
2° Protein-level
(e.g. drugs)
2° Protein-level
(e.g. drugs)
2° DNA/RNA-level
(e.g. siRNA)
2° Various growth
2° Various
Nature Reviews | Genetics
Figure 1 | Multidimensional genetic perturbation screens. a | The formats ( pooled
versus arrayed) and readouts routinely used in genetic perturbation screens are shown.
b | The graph depicts the cellular resolution, throughput and quantitative
information for different formats and readouts in genetic perturbation screens.
c | Combinatorial possibilities for genetic perturbation screens and the inference of
functional interactions from multidimensional data sets are shown. Depending on the two
types of perturbation the system is subject to, different type of genetic interactions can
be inferred. A third dimension that can be added to all previous systems is the single-cell
dimension. CRISPR–Cas9, clustered regularly interspaced short palindromic repeat–
CRISPR-associated protein 9; CyTOF, mass cytometry; FACS, fluorescence-activated cell
sorting; miRNA, microRNA; shRNA, short hairpin RNA; siRNA, small interfering RNA.
© 2014 Macmillan Publishers Limited. All rights reserved
Computer vision
A field that processes, analyses
and interprets images in order
to produce numerical
The local environment of a
single cell within a population
and their relative positioning to
each other, such as the local
crowding of cells, the amount
of neighbours, whether cells
face empty space on one site
and cells on another site, and
whether cells are solitary.
Cell segmentation
Automated detection and
delineation of the outside
of single cells and nuclei in
microscope images.
Cellular states
A quantitative description
of the physiological states of
single cells reflected in, for
instance, their sizes, shapes,
cell cycle phases, senescence
or other detectable readouts
such as metabolic states.
Gaussian mixture models
(GMMs). Parametric probability
density functions that fit
the Gaussian distribution
to the data set to model
the presence of
subpopulations within
an overall population;
they are represented as
weighted sums of Gaussian
component densities.
Support vector machine
(SVM). Supervised learning
models that recognize patterns
in data sets and that are used
for classification and regression
of an embryo or tissue, as well as in the ability to extract
a large number of quantitative features from segmented
cells and subcellular objects72,73 (FIG.2). This approach is
also referred to as ‘computer vision. Although imaging
lags behind flow cytometry in the number of molecular
readouts that can be simultaneously measured in single
cells, novel approaches74–76 indicate that imaging may
soon be able to achieve a similar or greater extent of
molecular multiplexing in single cells with an unprec-
edented subcellular resolution (BOX2). These techno-
logical advances in imaging can create vast amounts of
data from which, besides quantifying the abundance
of molecules or their activated (for example, phospho-
rylated) forms in single cells, much additional informa-
tion can be obtained on the subcellular localization and
patterning of the molecular signals77,78. Imaging can also
provide information on the morphology and shape of
single cells12,15 and, importantly, on the relative location
of a single cell with respect to other cells in a popula-
tion, as well as on its microenvironment, such as bordering
space without cells, the number of neighbours and local
crowding79 (FIG.2). For the purpose of this Review, we
further focus on image-based screens because this is the
only technique that allows the analysis of the full spec-
trum of cell-to-cell variability and that has the highest
spatial resolution (FIG.1b).
Managing multivariate single-cell readouts. A crucial
first step in arriving at high-quality multivariate single-
cell information is eliminating technical artefacts. In
imaging, this can be readily performed using approaches
from artificial intelligence, such as machine learning80,81.
Cell segmentation results can be evaluated using software
tools that project segmentation outlines on the cells in
images and, guided by this, classifiers can be trained to
identify incorrectly segmented cells, out-of-focus cells,
cells on image edges, cells with staining artefacts or, if
necessary, typical cellular states that produce outlier
values, such as mitotic and apoptotic cells80,81. It is not
uncommon that this process removes up to one-third of
all initially identified cells in a large-scale genetic pertur-
bation screen16. This exclusion step is important because
many statistical approaches using single-cell data, such
as principal component analysis (PCA) or Gaussian
mixture models (GMMs), are sensitive to outlier values
produced by such artefacts.
The second step to consider is the dimensionality of
the single-cell measurements (also known as features)
(see Supplementary information S1 (table)). Many
features correlate to such an extent that they basically
hold identical information. Therefore, a step of data
dimensionality reduction or feature elimination can
often be applied to reduce the data complexity with-
out losing information. Both linear methods (such as
PCA) and nonlinear methods (such as t-distributed
stochastic neighbour embedding (t-SNE)82,83) transform
the multi dimensional feature space into a space repre-
sented by fewer dimensions, which may allow an easier
computational handling of the data84,85. As transformed
dimensions can be difficult to interpret, an alternative
approach is to perform iterative feature elimination
without transforming the original feature space85. This
approach can reveal features that introduce unwanted
noise and thus weaken the computational identification
of single-cell phenotypes, as well as uncover features that
carry most of the information and thus strengthen such
Finally, it is often desired to perform classification of
single-cell phenotypes, such as changes in cellular mor-
phology12,15, perturbations in cellular activities4,11,16,43
or differences in patterns of intracellular organelles27,46.
Typically, either unsupervised machine learning
approaches (such as k-means clustering) or supervised
machine learning approaches (such as support vector
machine (SVM)) are used85 (see Supplementary informa-
tion S1 (table)). Supervised classification can be done
by browsing through the images and recognizing, by
eye, representative cells of different phenotypic classes,
which are then used to train a machine learning classi-
fier; alternatively, it can be done by using perturbations
that result in a known single-cell phenotype and classify-
ing all cells in the data set according to a set of known
perturbations4,12,15,46,86. Although this usually results in
great interpretability of the data, it has limitations. First,
it assumes that each cell must belong to a predefined
set of discrete classes and not part of a pheno typic con-
tinuum, which is particularly problematic for single cells
that are on the boundary of two or more classes. Second,
supervised approaches are not comprehensive because
they overlook single-cell phenotypes that are not apriori
known or expected. To address this issue, one can use
unsupervised approaches that classify all single cells in a
data set in an unbiased manner, such as single-cell clus-
tering or GMM87–91. The challenge with these unsuper-
vised approaches is that the single-cell clusters may not
always be biologically interpretable and are less robust
to ‘noisy’ features.
Both supervised and unsupervised classification
approaches assume that the single-cell phenotypic
space can be discretized. There is a natural inclination
to discretize data because it greatly aids interpretability.
Well-known concepts in genetics such as canalization
suggest that discrete phenotypes may emerge, and bista-
ble or multistable properties of some biological systems
also support a discretized view15,92. However, discreti-
zation of high-quality single-cell data can also lead to
a great loss of information and should, in our opinion,
be used with care. Some properties of single cells may
seem to be discrete in certain measurements, for exam-
ple, being in the G1 or G2 phase of the cell cycle when
measuring DNA content. However, the cell cycle repre-
sents a cycling pheno typic continuum in which certain
transitions occur faster than others, and the cell-to-cell
variability within each phase can be used as a proxy for
the time spent in that phase93. Moreover, many single-
cell measurements do not show discrete peaks within
a distribution, such as single-cell endocytic activity16.
By considering single cells on a phenotypic continuum,
one opens up to the concept of valuing the whole spec-
trum of cell-to-cell variability as biologically meaning-
ful. This concept is important when interpreting genetic
perturbation effects (see below).
© 2014 Macmillan Publishers Limited. All rights reserved
Common factor
Linear models
Smirnov test
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
497 187 406 781 367 645
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
1458 218 987 2896 385 145
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
521 964 7 146 789 175
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
227 415 254 516 577 1476
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
745 834 623 212 905 5
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
112 945 1367 631 461 234
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
487 510 190 762 845 111
Single cells
Gene 1
Gene 2
Gene 3
Gene 4
Gene n
Cells Population
1294 378 1225 804 748 846
729 1311 757 1827 247 954
1227 54 1497 191 1511 737
706 1512 130 1263 1408 1265
1455 1355 1074 1442 1095 1577
894 1686 1050 383 393 1216
113 905 976 1009 972 1183
1853 345 1750 852 457 113
1473 806 306 1974 1134 1361
682 1681 1256 948 1821 1722
1710 1708 702 498 954 1688
418 1698 976 734 613 53
352 760 1057 1369 455 694
1199 1915 665 49 276 1853
462 1993 779 1843 1156 1020
481 355 1147 917 910 897
806 1008 626 877 1708 712
1211 173 1128 1638 157 1665
1793 1502 923 1326 452 1983
997 1448 1172 1193 1872 865
257 1350 1909 694 4 1668
1773 216 1225 158 925 757
754 587 738 718 1892 536
1370 874 551 616 485 1267
227 415 254 516 577 1476
Phenotype and gene scoring
(e.g. k-means clustering)
Single cells
Nature Reviews | Genetics
Class 2Class 1
Number of neighbours
Local crowding Position islet Cellular sociology Population context
Class 1
Class 2
Class 3
Class 4
Class 5
(e.g. support vector machine)
Size and shape
Cell size
and shape
Protein content
Size and shape
DNA content
Cell cycle
Cell population
compartments CellNucleus Single cell
Figure 2 | Single-cell measurements and phenotypic scoring.
a | The different types of measurements (features) that can be
extracted from single cells and cell populations in images are
illustrated. b | Possible dimensionality in multivariate data sets (left
panel) is shown. Supervised and unsupervised methods can be used for
dimensionality reduction and for scoring single-cell phenotypes and
gene perturbations (right panel). Kullback–Leibler information
divergence is a non-symmetrical measure that calculates the difference
between two probability distributions. PCA, principal component
© 2014 Macmillan Publishers Limited. All rights reserved
Kolmogorov–Smirnov test
A statistical non‑parametric
test for the comparison of
continuous, one‑dimensional
probability distributions.
Population context
A collective term for the
context in which a single cell
displays its activities and
behaviours, which can be
determined by both local
and global effects from the
population to which the cell
belongs. The context is
determined not only by the
microenvironment of a single
cell but also by its physiological
state that is a consequence of
population effects, such as the
cell size.
The cell population context. Genetically identical cells
from the same population cultured in the same medium
can display a large spectrum of differences in pheno-
types and activity94–97 (FIG.3a). In fact, the variability in
single-cell properties within the same cell population
can be larger than the difference of the mean of these
activities between an unperturbed and a perturbed cell
population98. Therefore, in pioneering work, the classical
Kolmogorov –Smirnov test was used to compare single-cell
feature distributions11 (see Supplementary informa-
tionS1 (table)). By performing such tests on multiple
single-cell features separately for multiple-compound
treatments across a range of different concentrations,
this improved the ability to characterize drug responses
and to assign mechanisms to uncharacterized drugs11.
Although comparing full distributions instead of
mean values offers a powerful approach, its greatest limi-
tation is that it is not possible to determine which single
cells are compared with each other. Cell-to-cell variability
of different activities is determined by the physiologi-
cal and morphological state of a cell (that is, the cellular
state), which can be influenced by the microenvironment
in a non-trivial manner16,79,99105 (FIG.3a). As two single
cells that are in different microenvironments and cellular
states can both contribute to defining the same point in
a one-dimensional distribution, perturbations that act in
cells with different microenvironments or cellular states
can have a similar effect on this distribution. We illustrate
this for three theoretical single-cell features: the amount
of an organelle such as the lysosome, the amount of a
protein such as epithelial growth factor receptor (EGFR)
and the amount of the lipid GM1 at the plasma mem-
brane(FIG.3a). In each scenario, we show what happens
during perturbations if these features are either higher
in cells that grow at high local crowding, or higher in
big and spread-out cells that usually grow in regions of
low local crowding. Comparing only one-dimensional
single-cell distributions of activities would not distin-
guish between perturbations that alter the local crowd-
ing of cells or between perturbations that directly affect
the specific single-cell features independently of possible
effects on the cellular microenvironment16 (FIG.3b). To
take this into account, one should compare multivariate
single-cell distributions that incorporate features of the
cellular state and the cellular microenvironment1316. This
allows one to distinguish a situation in which a single-cell
activity distribution changed because the fraction of cells
in certain microenvironments changed (that is, indirect
perturbation) from a situation in which the single-cell
activity was directly perturbed (that is, direct perturba-
tion) (FIG.3b). Taking patterns of cell-to-cell variability
into account has been shown to be important for the
interpretation of genetic perturbations of virus infection,
endocytosis, membrane lipid composition and cell adhe-
sion signalling14,16,79. Furthermore, it is likely to be impor-
tant for studying other signal transduction pathways94,97
and thus for studying gene transcription, protein trans-
lation, metabolic activity and possibly also intrinsically
cycling cellular states such as the cell cycle106.
In large-scale genetic perturbation screens, many
perturbations can have effects on one or more aspects
of the cell population context14,16. These effects cannot be
predicted from only knowing the number of cells in a
population because of the nonlinear emergence of the
spectrum of single-cell microenvironments and states
in a growing population of cells (BOX1). Furthermore,
some perturbations might affect different properties of
the single-cell microenvironment and cellular state with-
out affecting the number of cells, for example, by altering
cell migration14,16 (BOX1). This possibility poses a seri-
ous problem for the comparison of genetic perturbation
effects and cannot be addressed with a correction for
trends in the single-cell readout as a function of cell num-
ber alone. Recently, computational approaches have been
developed that allow a more direct comparison of genetic
perturbation effects, provided that enough single cells are
quantified for each perturbation in a large-scale genetic
perturbation screen to permit statistical modelling14,16.
Finally, cellular heterogeneity and the fact that
genetic perturbations might affect different subsets of
cells within one population could have important con-
sequences for interpreting synthetic (that is, epistatic)
effects between two perturbations. The complexity of
this becomes very large if one takes into account that
one perturbation may indirectly alter population con-
text properties of single cells, which may allow another
perturbation to exert an effect. However, such synthetic
effects require a fundamentally different interpretation
compared with current practice, which usually assumes
that the targeted proteins are part of the same molecular
complex or pathway acting in the same single cells. This
population context effect may provide an explanation
for some of the problematic complexity encountered in
the analysis of multidimensional synthetic RNAi screens
in human cells64.
Box 2 | Multiplexing molecular readouts in imaging
In recent years, different approaches have been developed that allow the multiplexing
of molecular readouts in single cells in imaging74–76,152–154. One approach is based on
combinatorial labelling and spectral unmixing. This was applied in fluorescence insitu
hybridization (FISH) on a complex population of multiple bacterial species, using
simultaneous staining with 28 FISH probes that uniquely identified each species in the
population76. In another approach, also using FISH, spatial barcoding of long probes
with combinations of different fluorophores was applied, which could be identified
using super‑resolution microscopy. This approach allowed the multiplexing of 32
transcript readouts in single yeast cells154. It is conceivable that similar approaches can
be developed for antibodies. However, it is likely that these methods of multiplexing
will run into difficulties when the signals of multiple probes or antibodies overlap
within the same single cells. These problems do not arise with an alternative principle
for multiplexing, which relies on iterative staining and removal. In one approach, the
fluorescence signal of one stain is photobleached, after which a second stain can be
applied74. However, this becomes impractical when large numbers of single cells are
imaged. Other approaches rely on antibody elution, in which both primary and
secondary antibodies are removed by the use of detergent and low pH75,155, or on the
cleavage of oligonucleotides attached to antibodies152. Thus, it seems likely that, in
the near future, these developments will allow multiplexing of up to 100 molecular
readouts from thousands of single cells in an image‑based approach. The information
on the molecular state of single cells that is gained through this — combined with the
power of quantifying numerous morphological, spatial and patterning features, both
within single cells and across single cells — will empower multiple systems‑biology
approaches that no other method currently achieves. This would bridge the gap
between omics and single‑cell imaging.
© 2014 Macmillan Publishers Limited. All rights reserved
Overlay2 EGFR 3 GM1
Predicted activity per cell
Nature Reviews | Genetics
Cell countCell countCell count
1 Lysosomes
Activity 3Activity 1 Activity 2
b Activity 1: lysosomes
Activity Activity
Activity Activity
Activity Activity
Direct perturbation
Masked perturbation
Indirect perturbation
Measured activityMeasured activity Corrected activity
(relative to expected)
Expected activity
Expected distribution if the activity would
be unperturbed in the targeted population
Measured distribution in a control population
Measured distribution in the targeted population
Population context
Local cell density
Population size
Single-cell activity
Figure 3 | Accounting for population context in the interpretation
of genetic perturbation screens. a | Three theoretical single-cell
features are shown: the amount of an organelle such as the lysosome,
the amount of epithelial growth factor receptor (EGFR) at the
plasma membrane and the amount of the lipid GM1 at the plasma
membrane. Colour-coding on nuclear segmentation shows different
and non-trivial patterns of cell-to-cell variability in a cell population.
Three-dimensional (3D) surface plots below the images show predictors
of the cell-to-cell variability patterns of the three cellular activities.
b | Genetic perturbations can affect the population context of
individual cells (for example, the crowdedness of cells in a population)
and can directly or indirectly alter cellular activities. c | The different
ways by which a genetic perturbation can affect single-cell activities
are depicted. These can be mediated through indirect effects (via
changes in population context) or through direct effects on the
intracellular activity.
© 2014 Macmillan Publishers Limited. All rights reserved
Parallel phenotypic screens
Screens performed in parallel
in the same cell line using
the same perturbations but
different phenotypic readouts.
Accounting for the cell population context. The multi-
variate single-cell feature space of microenvironment
and cellular state can often be used to predict single-cell
activities in a non-parametric way14,16,79. For example, the
amount of intracellular organelles of a single cell can be
accurately predicted by the local crowding of the sin-
gle cell14,16. Such predictive models can be learned from
quantifying the cellular activity of interest in a large
number of unperturbed cells plated over a wide range
of cell numbers in a multiwell plate, and by creating
quantile bins of cells within the multi dimensional space
describing microenvironment and cellular state14,16.
For each perturbed cell population, such models pre-
dict what the expected single-cell activities, given their
specific microenvironment and cellular state, are. This
expected value is then subtracted from the measured
single-cell activities. By correcting for this effect, it
becomes possible to directly compare between two cell
populations from the same cell line (and also, to a certain
extent, from a different cell line), even if these cell popu-
lations show different distributions of single-cell micro-
environments and cellular states (FIG.3b). These effects
have a large impact on the interpretation of results from
a large-scale genetic perturbation screen14,16. Eliminating
the effects that act through population context phenom-
ena while maintaining single-cell autonomous effects
leads to a greater enrichment of genes encoding direct
regulatory and core machinery components of the cellu-
lar activity that is screened for14,16. In our three examples
of cellular activityabove (FIG.3a), which show different
patterns of cell-to-cell variability, indirect perturbation
effects determined by population context could act in
opposite directions on these readouts and could mask
direct effects16. Without correction, these types of effects
would dominate the interpretation of results, hiding
core machinery that is potentially shared. Importantly,
both types of effects are informative, and statistical
approaches to separate them are merely a means to
obtain more insights. Furthermore, such separations
have their limitations, particularly when there is exten-
sive and rapid feedback. However, as long as the direct
and indirect effects caused by the same genetic pertur-
bation show an additive effect and a large part of the
spectrum of cellular microenvironments is present in
the perturbed population, statistical approaches should
be able to separate them. Finally, population context
effects also occur in cell colonies of bacteria and yeast
grown on agar, suggesting that such analysis may also
be relevant for genetic perturbation screens in single-cell
Inferring functional interactions
To many researchers, the goal of large-scale genetic
perturbation screens is to identify groups of function-
ally related genes and gene interactions. However,
unbiased data-driven modelling on quantitative multi-
variate measurements from thousands of single cells
at a large scale now allows many types of interactions
to be identified from screens, revealing properties of
the studied cellular activity at multiple levels. This is a
systems-biology approach and goes beyond the simple
creation of hit lists to identify gene candidates for further
independent characterization. We describe below recent
developments in the identification and interpretation of
functional interactions between genes from perturbation
screens. Finally, we discuss a different type of interac-
tion that can be analysed from such screens — namely,
between systems properties (for example, single-cell
features) and perturbations.
Functional interactions between genes. The types of
gene–gene interactions that can be identified from
genetic perturbation screens are manifold. Here, we refer
to functional interaction as a global term that incorpo-
rates both physical interactions between two protein
subunits of the same complex (protein–protein inter-
actions) and classical genetic interactions measured by
epistasis between their respective loss-of-function effects
(FIG.4a). In our terminology, functional interactions
also include regulatory interactions16 such as kinase–
substrate interactions and phenotypic interactions,
which reflect the contribution of two genes to the emer-
gence of a specific phenotype, without any direct physical
or chemical interaction between the proteins (FIG.4a).
The inference of functional interactions and, more
specifically, classical genetic interactions has a long
history. The current ‘gold standard’ in mapping classi-
cal genetic interactions at a large scale is by means of
double-gene deletions, as these experimentally reveal
epistatic effects between two gene perturbations112,113.
The initial studies, pioneered in yeast, built a functional
genetic interaction map based on the measured epi-
static effects between two genes. However, it turned out
that the pairwise correlation between two genes across
a large set of epistatic effects with other genes is more
informative for predicting functionally related genes60,114.
Thus, statistical inference is used to derive an interac-
tion between two genes from the similarity in their
perturbation effects across a large number of genetic
backgrounds (which are caused by the second gene
perturbation). This concept is, in principle, not differ-
ent from inferring functional genetic interactions from
multiple single-gene perturbation screens performed in
parallel using either multiple environmental conditions
or multiple readouts16,115. Therefore, the large amount
of information that can be extracted from single-gene
perturbation screens using multivariate readouts of
thousands of single cells per perturbation could allow
the inference of genetic interactions of similar predictive
power as inferred genetic interactions from double-gene
perturbation screens16. In addition, as genetic interac-
tions are readout-dependent (see below), the interaction
inferred from parallel phenotypic screens may reveal novel
biology that cannot be revealed by screens based on
epistatic effects on colony fitness.
Genetic interactions are readout-dependent and plastic.
It is important to realize that the readout used in large-
scale genetic perturbation screens determines, to a large
extent, the functional interactions that one finds and the
novel biology that one uncovers64,116. Measuring colony
fitness reveals interactions between genes that severely
© 2014 Macmillan Publishers Limited. All rights reserved
Figure 4 | Inferring different types of genetic interactions.
a | Statistical inference of genetic interactions is shown. The schematic
shows the mode of action of overall correlations (yellow) to infer classical
genetic interactions versus subset effects in the data (for example, the
Hierarchical Interaction Score (HIS); blue) to infer regulatory interactions
between genes. b | Functional annotation enrichments in genes connected
by the HIS are compared to those in genes connected by overall correlation
inferred from a large double-gene perturbation screen in
Saccharomycescerevisiae61. Part b reprinted from Cell, 157, Liberali,P.,
Snijder,B. & Pelkmans,L., A hierarchical map of regulatory genetic
interactions in membrane trafficking, 1473–1487, Copyright (2014), with
permission from Elsevier.
Cell wall
Cell cycle
Vesicle transport
Aromatic amino acid
ATP binding
Protein import
RNA splicing
RNA degradation
Cellular activity of interest or variability
Low High
Nature Reviews | Genetics
Good correlation
Gene A Gene A
Gene B
Gene C
Phenotypic readouts or
epistatic readouts
DNA repair
385 genes6 genes
ATP binding
Membane traffic
DNA repair
Protein import
Spindle organization
RNA degradation
© 2014 Macmillan Publishers Limited. All rights reserved
The phenomenon whereby the
effects of a perturbation are a
subset of the effects of another
Hierarchical Interaction
(HIS). A statistical method that
infers functional interactions
between genes if they display
perturbation effects in a
consistent subset of readouts,
or environmental or genetic
backgrounds. It also infers
statistical hierarchy, in which
the perturbation with a
broader set of effects is placed
upstream of a perturbation
with a narrower subset of these
affect cell viability when perturbed61, whereas measur-
ing endocytosis will reveal interactions between genes
that affect those processes when perturbed16. With the
exception of a few protein complexes that are essential
for any measurable feature of cells (such as ribosomes
and proteasomes), these screens will reveal different
subsets of relevant functional interactions. Furthermore,
functional genetic interactions are highly plastic and can
show substantial differences when cells are exposed to
a single well-defined chemical perturbation117120. Thus,
there seems to be no single ground truth for functional
genetic interaction maps, which reflects the plasticity of
the underlying molecular networks.
Benchmarking functional genetic interactions. A wide
variety of different omic approaches are currently being
applied in numerous laboratories, leading to an explo-
sive growth in information on the molecular networks
of cells. This progress is being captured by numerous
databases that contain various types of information
about functional associations between genes, including
co-expression from microarrays, protein–protein inter-
actions and manually curated pathways121123. As a result,
there is an increasing tendency to compare results from
a large-scale genetic perturbation screen with such data-
bases, and computational approaches are developed to
use such databases as apriori information in the analy-
ses of large-scale genetic perturbation screens124,125. For
example, iterative feature selection can be applied to
compare the clustering of gene perturbations against
such databases, selecting or scaling features to improve
overlap126. However, the danger is that one could bias
the results of a screen to reveal expected interactions by
overfitting on the apriori data or by missing strong novel
interactions arising from the data. Furthermore, these
databases often generalize information and typically lack
contextual and dynamic information on interactions,
and are therefore currently far from being comprehen-
sive enough to be a useful apriori source of information,
especially when applied to areas of cell biology that do
not immediately link to classical systematic readouts
such as colony fitness or cell proliferation. However, we
consider the various databases of omic information to
be useful in determining general benchmarks for the
predictive power of unbiased statistical methods.
Statistical inference of genetic interactions. In recent
years, various statistical methods have been developed
and used for the analysis of double-gene perturbation
screens, and a broad set of modelling approaches have
been developed and applied to such data127129. This
includes, for example, ordinary differential equations
to model the final fitness measurements as the result of
a dynamic growth process130. However, the vast major-
ity of studies still rely on original clustering approaches,
which are based on the pairwise correlations between
two multivariate readouts of single-gene perturbations
or epistatic effects. Clustering comes from the field of
gene expression profiling, in which the clustering
of genes with similar transcript abundance patterns
in cells grown in different environmental or genetic
backgrounds allows the grouping of genes with similar
function131,132. Currently, clustering approaches using
perturbation data sets are often combined with orthogo-
nal types of large-scale data sets, For example, in an early
study functional modules were identified from the com-
bined evidence from the clustering of RNAi perturbation
effects on early cell division in the C.elegans embr yo,
protein–protein interactions identified using yeast two-
hybrid screens and clusters of mRNA expression pro-
files133. This approach mainly revealed well-known and
tightly interconnected molecular complexes, such as the
ribosome, the proteasome, the COPI coatomer, vacuolar
H+ ATPase and the anaphase-promoting complex133.
The use of overall correlations in clustering is based
on the assumption that two genes that co-function
within the same molecular complex, biosynthetic path-
way or regulatory mechanism must show the same per-
turbation effect (or synthetic effect) in many different
conditions or in many different readouts. Although this
is certainly true for many of the well-known interactions
(such as those within the ribosome or proteasome), this
is not a general rule. Many molecular complexes are
known to be dynamic and show different composi-
tions depending on the cellular activity in which they
are involved, or depending on the (micro)environmen-
tal context and intrinsic physiological state of the cells
in which they are acting. Furthermore, particularly in
regulatory interactions, an upstream kinase often phos-
phorylates numerous proteins involved in different cel-
lular activities. Therefore, perturbing this kinase might
result in a broad set of effects across several different
environmental or genetic backgrounds, multiple sin-
gle cells in a population or multiple different readouts.
Perturbing a downstream target of this kinase might
only share these effects for a subset of the conditions or
readouts and would thus demonstrate an overall poor
correlation with the kinase113,134. Vice versa, it might be
that a defined molecular complex is involved in multiple
activities but is regulated by different upstream kinases
in each of the activities. In this case, perturbing the
molecular machinery would have a broader set of effects
than perturbing one of the kinases. Moreover, a core set
of molecular machinery components may use different
subunits depending on the activity that the molecular
machinery is involved in. Such combinatorial use of
genes in cellular activities displays itself in large-scale
genetic perturbation screens in a phenomenon called
nesting, in which the effects of one perturbation are a
subset within the effects of an upstream perturbation135
(FIG.4b). Several methods have been developed to ana-
lyse such subset or nested effects136,137, but only recently
have they been advanced to be applicable to large-scale
genetic perturbation screens16,138. To capture these sub-
set effects, the Hierarchical Interaction Score (HIS)138 was
developed, which connects two perturbations if they
share a significant perturbation phenotype in at least
one of a set of multivariate readouts (FIG.4b). Moreover,
the HIS includes directionality in the connection if one
gene perturbation has a larger set of phenotypes than
the other gene perturbation (with the upstream one hav-
ing a larger set of phenotypes). When applied to a set of
© 2014 Macmillan Publishers Limited. All rights reserved
13 parallel RNAi screens measuring various aspects of
endocytosis16, this allowed the inference of regulatory
genetic interactions between genes involved in signal
transduction and membrane trafficking, and uncov-
ered novel regulatory connections within the endosomal
membrane system in humancells.
Interestingly, a comparison of functional genetic
interactions inferred by the HIS and by Pearson correla-
tion on the largest available collection of double-gene
perturbations performed in yeast (Saccharomycescerevi-
siae)58 revealed that both methods infer complementary
sets of functional interactions16,61. The HIS infers more
interactions between genes that are enriched in vesicle
transport, the endoplasmic reticulum and the Golgi, ATP
binding (mainly kinases), DNA repair, aromatic amino
acid biosynthesis, purine metabolism, ribonucleoprotein
complexes and RNA splicing. By contrast, correlations
mainly link genes that are enriched in transcription,
histone modification, cell wall biosynthesis, peroxisome
biogenesis, spindle localization and RNA degradation16
(FIG.4b). Apparently, some cellular processes rely more on
combinatorial or hierarchical genetic interactions than
others. As previous efforts in functional genetic interac-
tion mapping have overlooked hierarchical interactions,
the HIS provides an important additional approach to
study the functional genetic landscape ofcells.
Interactions between systems properties and perturba-
tions. Besides functional interactions between genes,
large-scale perturbation data sets can also be used
to analyse overall trends in the data or to generalize
whether certain features are co-perturbed in a multi-
variate feature set. Although machine learning, data
dimensionality reduction and feature elimination are
usually considered as data-processing steps, they do in
fact provide a first overview of exactly such trends. They
identify the features that correlate with each other across
a large number of perturbations and show that certain
combinations of feature values are more enriched in the
data set than others. In addition, multiple forms of data-
driven modelling — including partial correlation analy-
sis, multivariate regression or Bayesian network learning
— can be applied to multivariate single-cell measure-
ments to reveal systems properties of a cellular activity
and its response to genetic perturbations12,14,45,79,88,139.
When data-driven modelling is applied across a large
number of perturbations without a single-cell view (for
example, by single-cell averaging the multiple features),
this can reveal how certain properties of the cellular sys-
tem under investigation are more often co-perturbed
than other properties, which might indicate a causal
chain of events45. Importantly, such systems properties
can actually be revealed without the need for any per-
turbation by harnessing cell-to-cell variability — when
enough single cells are quantified, correlations and
causal interactions between properties can be inferred
from the variability present within one cell population.
If this cell-to-cell variability is combined with perturba-
tions, then such correlations and causal interactions can
become actual readouts in a screen, enabling the identi-
fication of genetic perturbations of systems properties.
For example, this approach allowed the identification
of genes perturbing the patterning of virus infection
in a cell population, without necessarily changing the
overall level of virus infection14. Without a single-cell
analysis, this effect would have remained unnoticed.
Such analyses are likely to become more predominant
in the future, as they aid the identification of functional
roles for genes in determining patterns of single-cell
activities across cell populations, allowing the emer-
gence of collective cellular behaviour, a fundamental
property of all lifeforms.
The ability to quantify multiple features of single cells
and to identify multiple subclasses of single-cell pheno-
types within a perturbed population has markedly
increased the statistical power with which gene pertur-
bations can be clustered13,15,70,140, but there are several
other implications that the single-cell paradigm might
have on the inference of functional genetic interactions.
Not only is correcting for population context-
determined effects important in the interpretation of
single-gene perturbation effects, but it may also aid
the interpretation of synthetic effects in double-gene
perturbations64,65. The common practice of correct-
ing for additive effects in synthetic interactions can
account for the situation where two perturbations
affect different subpopulation of cells, although it does
not account for indirect synergies between two per-
turbations through population context-determined
effects. One perturbation might change the micro-
environment of single cells, which makes these cells
sensitive to the second perturbation. Although such
interactions are interesting, they are different from a
direct functional interaction between two genes that
are part of, for example, the same signalling pathway
acting in the same single cell. The single-cell approach
might also offer some unique advantages for inferring
genetic interactions that have not yet been explored. If
we consider all the single cells in a population to repre-
sent (slightly) different phenotypic backgrounds, then
the comparison of two genetic perturbations across a
multivariate single-cell space could be just as predic-
tive of a functional interaction as comparisons across
a double-gene perturbation space without a single-cell
approach. This will require the quantification of a large
number of single cells in each perturbed population, as
well as a method that allows single cells with a similar
phenotypic background to be identified and compared
between two perturbed populations. The latter can
be achieved, as outlined, with features of the single-
cell microenvironment and state, which will become
even more powerful when combined with molecular
multiplexing (BOX2).
More general regulators of a cellular activity often
have a genetic perturbation effect in all cells of a popu-
lation, whereas other components might only show a
genetic perturbation effect in a subset of the population;
in this case, statistical methods that analyse subset effects
will be useful. Such population-determined subset
effects may thus be derived from a varying functional
© 2014 Macmillan Publishers Limited. All rights reserved
1. Grunenfelder,B. & Winzeler,E.A. Treasures and traps
in genome‑wide data sets: case examples from yeast.
Nature Rev. Genet. 3, 653–661 (2002).
2. Adams,M.D. & Sekelsky,J.J. From sequence to
phenotype: reverse genetics in Drosophila melanogaster.
Nature Rev. Genet. 3, 189–198 (2002).
3. Carpenter,A.E. & Sabatini,D.M. Systematic genome‑
wide screens of gene function. Nature Rev. Genet. 5,
11–22 (2004).
4. Loo,L.‑H., Wu,L.F. & Altschuler,S.J. Image‑based
multivariate profiling of drug responses from single
cells. Nature Methods 4, 445–453 (2007).
This paper identifies drug targets using a
SVM-based method that takes into account the
multivariate feature set of single cells in a
5. Boutros,M., Bras,L.P. & Huber,W. Analysis of
cell‑based RNAi screens. Genome Biol. 7, R66 (2006).
6. Boutros,M. & Ahringer,J. The art and design of
genetic screens: RNA interference. Nature Rev. Genet.
9, 554–566 (2008).
7. Birmingham,A. etal. Statistical methods for analysis
of high‑throughput RNA interference screens. Nature
Methods 6, 569–575 (2009).
8. Schadt,E.E., Linderman,M.D., Sorenson,J., Lee,L.
& Nolan,G.P. Computational solutions to large‑scale
data management and analysis. Nature Rev. Genet.
11, 647–657 (2010).
9. Friedman,A. & Perrimon,N. Genome‑wide high‑
throughput screens in functional genomics. Curr. Opin.
Genet. Dev. 14, 470–476 (2004).
10. Mohr,S., Bakal,C. & Perrimon,N. Genomic screening
with RNAi: results and challenges. Annu. Rev.
Biochem. 79, 37–64 (2010).
11. Perlman,Z.E. etal. Multidimensional drug profiling
by automated microscopy. Science 306, 1194–1198
This pioneering paper identifies drug targets by
taking into account the full distribution of single
cells in a population using the Kolmogorov–Smirnov
test in image-based small-compound screens.
12. Bakal,C., Aach,J., Church,G. & Perrimon,N.
Quantitative morphological signatures define local
signaling networks regulating cell morphology.
Science 316, 1753–1756 (2007).
This study uses quantitative morphological features
of single cells as profiles to identify genes involved
in cellular morphology.
13. Loo,L.‑H. etal. An approach for extensibly profiling
the molecular states of cellular subpopulations.
Nature Methods 6, 759–765 (2009).
14. Snijder,B. etal. Single‑cell analysis of population
context advances RNAi screening at multiple levels.
Mol. Systems Biol. 8, 579 (2012).
This paper shows that modelling single-cell
behaviour and taking into account cell-to-cell
variability strongly improve the data and
comparability in siRNA screens.
15. Yin,Z. etal. A screen for morphological complexity
identifies regulators of switch‑like transitions between
discrete cell shapes. Nature Cell Biol. 15, 860–871
16. Liberali,P., Snijder,B. & Pelkmans,L. A hierarchical
map of regulatory genetic interactions in membrane
trafficking. Cell 157, 1473–1487 (2014).
This paper infers regulatory genetic interactions
from parallel siRNA screens in human cells and
from double-knockout synthetic screens in yeast
using the HIS.
17. Novick,P., Field,C. & Schekman,R. Identification of
23 complementation groups required for post‑
translational events in the yeast secretory pathway.
Cell 21, 205–215 (1980).
18. Gonczy,P. etal. Functional genomic analysis of cell
division in C.elegans using RNAi of genes on
chromosome III. Nature 408, 331–336 (2000).
19. Fraser,A.G. etal. Functional genomic analysis of
C.elegans chromosome I by systematic RNA
interference. Nature 408, 325–330 (2000).
20. Lum,L. etal. Identification of Hedgehog pathway
components by RNAi in Drosophila cultured cells.
Science 299, 2039–2045 (2003).
21. Aza‑Blanc,P. etal. Identification of modulators of
TRAIL‑induced apoptosis via RNAi‑based phenotypic
screening. Mol. Cell 12, 627–637 (2003).
22. Brummelkamp,T.R., Nijman,S.M., Dirac,A.M. &
Bernards,R. Loss of the cylindromatosis tumour
suppressor inhibits apoptosis by activating NF‑κB.
Nature 424, 797–801 (2003).
23. Heo,W.D. & Meyer,T. Switch‑of‑function mutants
based on morphology classification of Ras superfamily
small GTPases. Cell 113 , 315–328 (2003).
24. Boutros,M. etal. Genome‑wide RNAi analysis of
growth and viability in Drosophila cells. Science 303,
832–835 (2004).
25. Paddison,P.J. etal. A resource for large‑scale RNA
interference‑based screens in mammals. Nature 428,
427–431 (2004).
26. Neumann,B. etal. Phenotypic profiling of the human
genome by time‑lapse microscopy reveals cell division
genes. Nature 464, 721–727 (2010).
This is the first genome-wide siRNA screen using
time-lapse imaging of living cells, which extracted
features from dynamic data to identify genes
involved in mitosis.
27. Simpson,J.C. etal. Genome‑wide RNAi screening
identifies human proteins with a regulatory function in
the early secretory pathway. Nature Cell Biol. 14,
764–774 (2012).
28. Stevenson,L.F., Kennedy,B.K. & Harlow,E.
A large‑scale overexpression screen in Saccharomyces
cerevisiae identifies previously uncharacterized
cell cycle genes. Proc. Natl Acad. Sci. USA 98,
3946–3951 (2001).
29. Pritsker,M., Ford,N.R., Jenq,H.T. & Lemischka,I.R.
Genomewide gain‑of‑function genetic screen identifies
functionally active genes in mouse embryonic stem
cells. Proc. Natl Acad. Sci. USA 103, 6946–6951
30. Stockwell,B.R. Chemical genetics: ligand‑based
discovery of gene function. Nature Rev. Genet.
1,116–125 (2000).
31. Durai,S. etal. Zinc finger nucleases: custom‑designed
molecular scissors for genome engineering of
plant and mammalian cells. Nucleic Acids Res. 33,
5978–5990 (2005).
32. Miller,J.C. etal. A TALE nuclease architecture
for efficient genome editing. Nature Biotech. 29,
143–148 (2011).
33. Urnov,F.D., Rebar,E.J., Holmes,M.C., Zhang,H.S.
& Gregory,P.D. Genome editing with engineered zinc
finger nucleases. Nature Rev. Genet. 11, 636–646
34. Cong,L. etal. Multiplex genome engineering using
CRISPR/Cas systems. Science 339, 819–823 (2013).
35. Mali,P. etal. RNA‑guided human genome engineering
via Cas9. Science 339, 823–826 (2013).
This paper shows the first proof of principle of the
CRISPR–Cas9 system in human cells.
36. Shalem,O. etal. Genome‑scale CRISPR–Cas9
knockout screening in human cells. Science 343,
84–87 (2014).
Using gene editing with the CRISPR–Cas9 system,
this paper establishes genome-scale gene
perturbation screening in a pooled format in
human cancer and pluripotent stem cells.
37. Wang,T., Wei,J.J., Sabatini,D.M. & Lander,E.S.
Genetic screens in human cells using the CRISPR–
Cas9 system. Science 343, 80–84 (2014).
Using gene editing with the CRISPR–Cas9 system,
this paper establishes genome-scale gene
perturbation screening in a pooled format in
haploid and diploid cell lines.
38. Hsu,P.D., Lander,E.S. & Zhang,F. Development and
applications of CRISPR–Cas9 for Genome engineering.
Cell 157, 1262–1278 (2014).
39. Qi,L.S. etal. Repurposing CRISPR as an RNA‑guided
platform for sequence‑specific control of gene
expression. Cell 152, 1173–1183 (2013).
40. Gilbert, L. A. et al. Genome‑scale CRISPR‑mediated
control of gene repression and activation. Cell 159,
647–661 (2014).
In references 39 and 40, the researchers
repurposed the CRISPR system to induce
sequence-specific repression (CRISPRi) or
activation (CRISPRa) of gene expression at a
genome scale.
41. Eggert,U.S., Field,C.M. & Mitchison,T.J.
Small molecules in an RNAi world. Mol. BioSystems 2,
93 (2006).
42. Weiss,W.A., Taylor,S.S. & Shokat,K.M.
Recognizing and exploiting differences between RNAi
and small‑molecule inhibitors. Nature Chem. Biol. 3,
739–744 (2007).
43. Bakal,C. etal. Phosphorylation networks regulating
JNK activity in diverse genetic backgrounds. Science
322, 453–456 (2008).
This paper reports a high-throughput screen that
uses RNAi to systematically inhibit two genes
simultaneously in 17,724 combinations to study
kinase regulation.
44. Pelkmans,L. etal. Genome‑wide analysis of human
kinases in clathrin‑ and caveolae/raft‑mediated
endocytosis. Nature 436, 78–86 (2005).
This paper is the first to report parallel
comparative siRNA screens.
45. Collinet,C. etal. Systems survey of endocytosis
by multiparametric image analysis. Nature 464,
243–249 (2010).
This paper reports the first multivariate
genome-wide screen of endocytosis.
46. Chia,J. etal. RNAi screening reveals a large signaling
network controlling the Golgi apparatus in human
cells. Mol. Systems Biol. 8, 1–33 (2012).
47. Silva,J.M. etal. Profiling essential genes in human
mammary cells by multiplex RNAi screening. Science
319, 617–620 (2008).
48. Bassik,M.C. etal. A systematic mammalian genetic
interaction map reveals pathways underlying ricin
susceptibility. Cell 152, 909–922 (2013).
In this study, the researchers construct a
double-shRNA library for pooled screens in
human cells to identify genetic interaction between
genes involved in ricin toxin susceptibility.
49. Carette,J.E. etal. Haploid genetic screens in human
cells identify host factors used by pathogens. Science
326, 1231–1235 (2009).
This is the first pooled screen in mammalian
haploid cells using random mutational insertions.
50. Winzeler,E.A. etal. Functional characterization
of the S.cerevisiae genome by gene deletion
and parallel analysis. Science 285, 901–906
51. Giaever,G. etal. Functional profiling of the
Saccharomyces cerevisiae genome. Nature 418,
387–391 (2002).
This paper reports the construction of a collection
of all viable single-gene deletion mutants of
S. cerevisiae.
engagement of genes in a cellular activity across single
cells in a population, and analyses of these patterns
could reveal parallel pathways and regulatory net-
works depending on the single-cell microenvironment
and state. Given that even classical functional genetic
interactions change in the presence of a single chemi-
cal perturbation118, it might be that different maps of
functional genetic interactions will also be found within
the same cell population depending on the single-cell
micro environment and state. Clearly, additional research
will be required to investigate all of these aspects. With
the single-cell paradigm now firmly established in biol-
ogy and with rapid technological developments, it will
soon be possible for research groups across disciplines to
take the phenotypic spectrum of single cells into account
in the interpretation of their experiments. This advance is
inevitable and essential — we expect that it will improve
interpretations of genetic perturbation screens and
provide a rich ground for numerous novel biological
discoveries for years tocome.
© 2014 Macmillan Publishers Limited. All rights reserved
52. Kim,D.U. etal. Analysis of a genome‑wide set of gene
deletions in the fission yeast Schizosaccharomyces
pombe. Nature Biotech. 28, 617–623 (2010).
53. Tang,T. etal. A mouse knockout library for secreted
and transmembrane proteins. Nature Biotech. 28,
749–755 (2010).
54. Dolgin,E. Mouse library set to be knockout. Nature
474, 262–263 (2011).
55. Skarnes,W.C. etal. A conditional knockout resource
for the genome‑wide study of mouse gene function.
Nature 474, 337–342 (2011).
56. Kotecki,M., Reddy,P.S. & Cochran,B.H.
Isolation and characterization of a near‑haploid
human cell line. Exp. Cell Res. 252, 273–280 (1999).
57. Leeb,M. & Wutz,A. Derivation of haploid embryonic
stem cells from mouse embryos. Nature 479,
131–134 (2011).
58. Burckstummer,T. etal. A reversible gene trap
collection empowers haploid genetics in human cells.
Nature Methods 10, 965–971 (2013).
59. Tong,A.H. etal. Systematic genetic analysis with
ordered arrays of yeast deletion mutants. Science
294, 2364–2368 (2001).
In this study, the researchers develop a method for
the systematic construction of a double-gene
deletion library for synthetic screens in
60. Schuldiner,M. etal. Exploration of the function and
organization of the yeast early secretory pathway
through an epistatic miniarray profile. Cell 123,
507–519 (2005).
This paper introduces the use of pairwise
correlations between two genes across a large set
of epistatic effects to infer functional genetic
interactions in S. cerevisiae.
61. Costanzo,M. etal. The genetic landscape of a cell.
Science 327, 425–431 (2010).
In this study, researchers create a genome-scale
genetic interaction map by examining 5.4 million
gene–gene pairs for synthetic genetic interactions
in S. cerevisiae.
62. Eggert,U.S. etal. Parallel chemical genetic and
genome‑wide RNAi screens identify cytokinesis
inhibitors and targets. PLoS Biol. 2, e379 (2004).
63. Jiang,H., Pritchard,J.R., Williams,R.T.,
Lauffenburger,D.A. & Hemann,M.T. A mammalian
functional‑genetic approach to characterizing cancer
therapeutics. Nature Chem. Biol. 7, 92–100 (2011).
64. Laufer,C., Fischer,B., Billmann,M., Huber,W. &
Boutros,M. Mapping genetic interactions in human
cancer cells with RNAi and multiparametric
phenotyping. Nature Methods 10, 427–431 (2013).
This paper shows an arrayed synthetic screen in
mammalian cells based on double-gene
perturbation with RNAi and uses multiple readouts
from single cells to infer genetic interactions.
65. Roguev,A. etal. Quantitative genetic‑interaction
mapping in mammalian cells. Nature Methods 10,
432–437 (2013).
This paper shows an arrayed synthetic screen in
mammalian cells based on double-gene perturbation
with RNAi to infer genetic interactions, which are
compared to protein–protein interaction data.
66. Reiling,J.H. etal. A haploid genetic screen identifies
the major facilitator domain containing 2A (MFSD2A)
transporter as a key mediator in the response
to tunicamycin. Proc. Natl Acad. Sci. USA 108,
11756–11765 (2011).
67. Barbie,D.A. etal. Systematic RNA interference
reveals that oncogenic KRAS‑driven cancers require
TBK1. Nature 462, 108–112 (2009).
This study uses siRNA screens to detect synthetic
lethal partners of oncogenic KRAS.
68. Ashworth,A., Lord,C.J. & Reis,J.S. Genetic
interactions in cancer progression and treatment.
Cell 145, 30–38 (2011).
69. Bandura,D.R. etal. Mass cytometry: technique for
real time single cell multitarget immunoassay based
on inductively coupled plasma time‑of‑flight mass
spectrometry. Anal. Chem. 81, 6813–6822 (2009).
70. Bendall,S.C. etal. Single‑cell mass cytometry of
differential immune and drug responses across a
human hematopoietic continuum. Science 332,
687–696 (2011).
71. Bodenmiller,B. etal. Multiplexed mass cytometry
profiling of cellular states perturbed by small‑
molecule regulators. Nature Biotech. 30, 858–867
72. Carpenter,A.E. etal. CellProfiler: image analysis
software for identifying and quantifying cell
phenotypes. Genome Biol. 7, R100 (2006).
73. Eliceiri,K.W. etal. Biological imaging software tools.
Nature Methods 9, 697–710 (2012).
74. Schubert,W. etal. Analyzing proteome topology and
function by automated multidimensional fluorescence
microscopy. Nature Biotech. 24, 1270–1278 (2006).
75. Zrazhevskiy,P. & Gao,X. Quantum dot imaging
platform for single‑cell molecular profiling. Nature
Commun. 4, 1619 (2013).
76. Valm,A.M. etal. Systems‑level analysis of microbial
community organization through combinatorial
labeling and spectral imaging. Proc. Natl Acad. Sci.
USA 108, 4152–4157 (2011).
77. Dehmelt,L. & Bastiaens,P.I. Spatial organization of
intracellular communication: insights from imaging.
Nature Rev. Mol. Cell Biol. 11 , 440–452 (2010).
78. Welch,C.M., Elliott,H., Danuser,G. & Hahn,K.M.
Imaging the coordination of multiple signalling
activities in living cells. Nature Rev. Mol. Cell Biol. 12,
749–756 (2011).
79. Snijder,B. etal. Population context determines
cell‑to‑cell variability in endocytosis and virus
infection. Nature 461, 520–523 (2009).
This paper shows that cell-to-cell variability in a
population of monoclonal cells is not stochastic but
can be predicted at the single-cell level by features
of the cellular state and microenvironment.
80. Ramo,P., Sacher,R., Snijder,B., Begemann,B. &
Pelkmans,L. CellClassifier: supervised learning of
cellular phenotypes. Bioinformatics 25, 3028–3030
81. Jones,T.R. etal. CellProfiler Analyst: data
exploration and analysis software for complex
image‑based screens. BMC Bioinformatics 9, 482
82. Bendall,S.C. etal. Single‑cell trajectory detection
uncovers progression and regulatory coordination in
human Bcell development. Cell 157, 714–725
83. Hinton, G. & van der Maaten, L. Visualizing data using
t‑SNE. J.Machine Learning Research 9, 2579–2605
84. Li,L. Dimension reduction for high‑dimensional data.
Methods Mol. Biol. 620, 417–434 (2010).
85. Duda,R.O., Hart,P.E. & Stork,D.G. Pattern
Classification (John Wiley, 2001).
86. Jones,T.R. etal. Scoring diverse cellular
morphologies in image‑based screens with iterative
feedback and machine learning. Proc. Natl Acad. Sci.
USA 106, 1826–1831 (2009).
87. Slack,M.D., Martinez,E.D., Wu,L.F. &
Altschuler,S.J. Characterizing heterogeneous cellular
responses to perturbations. Proc. Natl Acad. Sci. USA
105, 19306–193011 (2008).
88. Singh,D.K. etal. Patterns of basal signaling
heterogeneity can distinguish cellular populations with
different drug sensitivities. Mol. Syst. Biol. 6, 639
89. Zhong,Q., Busetto,A.G., Fededa,J.P.,
Buhmann,J.M. & Gerlich,D.W. Unsupervised
modeling of cell morphology dynamics for time‑
lapse microscopy. Nature Methods 9, 711–713
90. Qiu,P. etal. Extracting a cellular hierarchy from high‑
dimensional cytometry data with SPADE. Nature
Biotech. 29, 886–891 (2011).
91. Battich,N., Stoeger,T. & Pelkmans,L. Image‑based
transcriptomics in thousands of single human cells at
single‑molecule resolution. Nature Methods 10,
1127–1133 (2013).
92. Waddington,C.H. Canalization of development and
the inheritance of acquired characters. Nature 150,
563–565 (1942).
93. Kafri,R. etal. Dynamics extracted from fixed cells
reveal feedback linking cell growth to cell cycle. Nature
494, 480–483 (2013).
94. Spencer,S.L., Gaudet,S., Albeck,J.G., Burke,J.M. &
Sorger,P.K. Non‑genetic origins of cell‑to‑cell
variability in TRAIL‑induced apoptosis. Nature 459,
428–432 (2009).
95. Altschuler,S.J. & Wu,L.F. Cellular heterogeneity: do
differences make a difference? Cell 141, 559–563
96. Snijder,B. & Pelkmans,L. Origins of regulated
cell‑to‑cell variability. Nature Rev. Mol. Cell Biol. 12,
119–125 (2011).
97. Yuan,T.L., Wulf,G., Burga,L. & Cantley,L.C.
Cell‑to‑cell variability in PI3K protein level regulates
PI3K–AKT pathway activity in cell populations.
Curr. Biol. 21, 173–183 (2011).
98. Keren,K. etal. Mechanism of shape determination in
motile cells. Nature 453, 475–480 (2008).
99. Colman‑Lerner,A. etal. Regulated cell‑to‑cell variation
in a cell‑fate decision system. Nature 437, 699–706
100. Castor,L.N. Flattening, movement and control of
division of epithelial‑like cells. J.Cell. Physiol. 75,
57–64 (1970).
101. Eagle,H., Levine,E.M. & Koprowski,H. Species
specificity in growth regulatory effects of cellular
interaction. Nature 220, 266–269 (1968).
102. Eagle,H. & Levine,E.M. Growth regulatory effects of
cellular interaction. Nature 213, 1102–1106 (1967).
103. Zeng,L. etal. Decision making at a subcellular level
determines the outcome of bacteriophage infection.
Cell 141, 682–691 (2010).
104. St‑Pierre,F. & Endy,D. Determination of cell fate
selection during phage λ infection. Proc. Natl Acad.
Sci. USA 105, 20705–20710 (2008).
105. Robert,L. etal. Pre‑dispositions and epigenetic
inheritance in the Escherichia coli lactose operon
bistable switch. Mol. Syst. Biol. 6, 357 (2010).
106. Pelkmans,L. Cell Biology. Using cell‑to‑cell variability
— a new era in molecular biology. Science 336,
425–426 (2012).
107. Halme,A., Bumgarner,S., Styles,C. & Fink,G.R.
Genetic and epigenetic regulation of the FLO gene
family generates cell‑surface variation in yeast. Cell
116 , 405–415 (2004).
108. Avery,S.V. Microbial cell individuality and the
underlying sources of heterogeneity. Nature Rev.
Microbiol. 4, 577–587 (2006).
109. Vlamakis,H., Aguilar,C., Losick,R. & Kolter,R.
Control of cell fate by the formation of an
architecturally complex bacterial community. Genes
Dev. 22, 945–953 (2008).
110 . Parsons,B.D., Schindler,A., Evans,D.H. &
Foley,E.A. Direct phenotypic comparison of siRNA
pools and multiple individual duplexes in a functional
assay. PLoS ONE 4, e8471 (2009).
111. Cap,M., Stepanek,L., Harant,K., Vachova,L. &
Palkova,Z. Cell differentiation within a yeast colony:
metabolic and regulatory parallels with a tumor‑
affected organism. Mol. Cell 46, 436–448 (2012).
112 . Dixon,S.J., Costanzo,M., Baryshnikova,A.,
Andrews,B. & Boone,C. Systematic mapping of
genetic interaction networks. Annu. Rev. Genet. 43,
601–625 (2009).
113 . Collins,S.R., Roguev,A. & Krogan,N.J. Quantitative
Genetic Interaction Mapping Using the E-MAP
Approach Ch. 9 (Elsevier, 2010).
114 . Tong,A.H.Y. etal. Global mapping of the yeast genetic
interaction network. Science 303, 808–813 (2004).
115 . Nichols,R.J. etal. Phenotypic landscape of a bacterial
cell. Cell 144, 143–156 (2011).
This study combines large-scale chemical genomics
with quantitative fitness measurements in
hundreds of parallel conditions in Escherichiacoli.
116 . Horn,T. etal. Mapping of signaling networks through
synthetic genetic interaction analysis by RNAi. Nature
Methods 8, 341–U391 (2011).
117 . Guénolé,A. etal. Dissection of DNA damage
responses using multiconditional genetic interaction
maps. Mol. Cell 49, 346–358 (2013).
118 . Bandyopadhyay,S. etal. Rewiring of genetic
networks in response to DNA damage. Science 330,
1385–1389 (2010).
119 . Ideker,T. & Krogan,N.J. Differential network biology.
Mol. Systems Biol. 8, 565 (2012).
120. Santos,S.D.M., Verveer,P.J. & Bastiaens,P.I.H.
Growth factor‑induced MAPK network topology
shapes Erk response determining PC‑12 cell fate.
Nature Cell Biol. 9, 324–U139 (2007).
121. Huang,D.W., Sherman,B.T. & Lempicki,R.A.
Systematic and integrative analysis of large gene lists
using DAVID bioinformatics resources. Nature Protoc.
4, 44–57 (2009).
122. Szklarczyk,D. etal. The STRING database in 2011:
functional interaction networks of proteins, globally
integrated and scored. Nucleic Acids Res. 39,
D561–D568 (2011).
123. Cerami,E.G. etal. Pathway Commons, a web resource
for biological pathway data. Nucleic Acids Res. 39,
D685–D690 (2011).
124. Barabasi,A.L., Gulbahce,N. & Loscalzo,J. Network
medicine: a network‑based approach to human
disease. Nature Rev. Genet. 12, 56–68 (2011).
125. Markowetz,F. How to understand the cell by breaking
it: network analysis of gene perturbation screens.
PLoS Comput. Biol. 6, e1000655 (2010).
126. Guyon,I. & Elisseeff,A. An introduction to variable
and feature selection. J.Machine Learn. Res. 3,
1157–1182 (2003).
© 2014 Macmillan Publishers Limited. All rights reserved
127. Wang,L., Wang,X., Arkin,A.P. & Samoilov,M.S.
Inference of gene regulatory networks from
genome‑wide knockout fitness data. Bioinformatics
29, 338–346 (2013).
128. Boone,C., Bussey,H. & Andrews,B.J. Exploring
genetic interactions and networks with yeast. Nature
Rev. Genet. 8, 437–449 (2007).
129. Battle,A., Jonikas,M.C., Walter,P., Weissman,J.S. &
Koller,D. Automated identification of pathways from
quantitative genetic interaction data. Mol. Syst. Biol.
6, 379 (2010).
130. Wang,L.M., Wang,X.D., Arkin,A.P. &
Samoilov,M.S. Inference of gene regulatory networks
from genome‑wide knockout fitness data.
Bioinformatics 29, 338–346 (2013).
131. Eisen,M.B., Spellman,P.T., Brown,P.O. &
Botstein,D. Cluster analysis and display of genome‑
wide expression patterns. Proc. Natl Acad. Sci. USA
95, 14863–14868 (1998).
132. Hughes,T.R. etal. Functional discovery via a
compendium of expression profiles. Cell 102,
109–126 (2000).
133. Gunsalus,K.C. etal. Predictive models of molecular
machines involved in Caenorhabditis elegans early
embryogenesis. Nature 436, 861–865 (2005).
134. Fiedler,D. etal. Functional organization of the
S.cerevisiae phosphorylation network. Cell 136,
952–963 (2009).
This paper analyses synthetic interactions between
gene knockouts of kinases, phosphatases and their
substrates in S. cerevisiae. It shows that kinases,
phosphatases and their substrates have positive
epistatic interactions between each other but no
significant correlation between their epistatic
effect profiles.
135. Boutros,M., Agaisse,H. & Perrimon,N. Sequential
activation of signaling pathways during innate immune
responses in Drosophila. Dev. Cell 3, 711–722
136. Markowetz,F., Bloch,J. & Spang,R.
Non‑transcriptional pathway features reconstructed
from secondary effects of RNA interference.
Bioinformatics 21, 4026–4032 (2005).
137. Markowetz,F., Kostka,D., Troyanskaya,O.G. &
Spang,R. Nested effects models for high‑dimensional
phenotyping screens. Bioinformatics 23, i305–i312
138. Snijder,B., Liberali,P., Frechin,M., Stoeger,T. &
Pelkmans,L. Predicting functional gene interactions
with the hierarchical interaction score. Nature
Methods 10, 1089–1092 (2013).
139. Young,D.W. etal. Integrating high‑content
screening and ligand–target prediction to identify
mechanism of action. Nature Chem. Biol. 4, 59–68
140. Irish,J.M. etal. Single cell profiling of potentiated
phospho‑protein networks in cancer cells. Cell 118,
217–228 (2004).
141. Elowitz,M.B. Stochastic gene expression in a
single cell. Science 297, 1183–1186 (2002).
142. Eldar,A. & Elowitz,M.B. Functional roles for noise
in genetic circuits. Nature 467, 167–173 (2010).
143. Raj,A., Rifkin,S.A., Andersen,E. & van
Oudenaarden,A. Variability in gene expression
underlies incomplete penetrance. Nature 463,
913–918 (2010).
144. Munsky,B., Neuert,G. & van Oudenaarden,A.
Using gene expression noise to understand gene
regulation. Science 336, 183–187 (2012).
145. Stelling,J. etal. Robustness of cellular functions.
Cell 118 , 675–685 (2004).
146. Macarthur,B.D., Ma’ayan,A. & Lemischka,I.R.
Systems biology of stem cell fate and cellular
reprogramming. Nature Rev. Mol. Cell Biol. 10,
672–681 (2009).
147. Barad,O. etal. Robust selection of sensory organ
precursors by the Notch‑δ pathway. Curr. Opin. Cell
Biol. 23, 663–667 (2011).
148. Ribrault,C., Sekimoto,K. & Triller,A. From the
stochasticity of molecular processes to the variability
of synaptic transmission. Nature Rev. Neurosci. 12,
375–387 (2011).
149. Brandman,O. & Meyer,T. Feedback loops shape
cellular signals in space and time. Science 322,
390–395 (2008).
150. Connelly,J.T. etal. Actin and serum response factor
transduce physical cues from the microenvironment to
regulate epidermal stem cell fate decisions. Nature
Cell Biol. 12, 711–718 (2010).
151. Engler,A.J., Sen,S., Sweeney,H.L. & Discher,D.E.
Matrix elasticity directs stem cell lineage specification.
Cell 126, 677–689 (2006).
152. Ullal,A.V. etal. Cancer cell profiling by
barcoding allows multiplexed protein analysis
in fine‑needle aspirates. Sci. Transl Med. 6,
219ra9 (2014).
153. Jungmann,R. etal. Multiplexed 3D cellular
super‑resolution imaging with DNA‑PAINT and
Exchange‑PAINT. Nature Methods 11 , 313–318
154. Lubeck,E. & Cai,L. Single‑cell systems biology by
super‑resolution imaging and combinatorial labeling.
Nature Methods 9, 743–748 (2012).
155. Gerdes,M.J. etal. Highly multiplexed single‑cell
analysis of formalin‑fixed, paraffin‑embedded cancer
tissue. Proc. Natl Acad. Sci. USA 110, 11982–11987
The authors thank M.Muellner and S.Nijman for the images
in the Box1, and all members of the Pelkmans laboratory for
discussions. P.L. is supported by a FEBS postdoctoral fellow
ship. B.S. is supported by an advanced postdoc fellowship of
SNSF. Research in the Pelkmans laboratory on these topics is
funded by the University of Zürich Research Priority Program
(URPP) in Functional Genomics and Systems Biology, and the
Swiss initiative in Systems Biology:
Competing interests statement
The authors declare no competing interests.
RNAi Database:
Saccharomyces Genome Database: http://www.
Systems survey of endocytosis by multiparametric image
See online article: S1 (table)
© 2014 Macmillan Publishers Limited. All rights reserved
... This kind of data can capture part of the complexity of the phenotypic state of cells and the impact of chemical or genetic perturbations therein, and can thus be very powerful to map gene networks and mechanisms involved in a given biological process (Liberali et al., 2014;Sailem and Bakal, 2017). Moreover, these studies can also capture cell-and populationcontext features, such as confluence, heterogeneity, spreading and morphology, or phenotypic signatures derived from other cell structures: information typically lost in other approaches (Liberali et al., 2015;Gut et al., 2018). Within-tissue heterogeneity in LD biogenesis, for example, is a striking but poorly understood phenomenon that could contribute to metabolic flexibility, and maximize protection from lipotoxicity (Herms et al., 2013). ...
... Functional genomics on specific human cell models have also been developed and applied to gain systems-level insights onto organelle biology (Liberali et al., 2014;Liberali et al., 2015). Importantly, these approaches can provide information regarding upstream signaling pathways and cell functions influencing a given phenotype. ...
Full-text available
Lipid droplets (LDs) are spherical, single sheet phospholipid-bound organelles that store neutral lipids in all eukaryotes and some prokaryotes. Initially conceived as relatively inert depots for energy and lipid precursors, these highly dynamic structures play active roles in homeostatic functions beyond metabolism, such as proteostasis and protein turnover, innate immunity and defense. A major share of the knowledge behind this paradigm shift has been enabled by the use of systematic molecular profiling approaches, capable of revealing and describing these non-intuitive systems-level relationships. Here, we discuss these advances and some of the challenges they entail, and highlight standing questions in the field.
... Routine comparative approaches included using Venn diagrams to identify hits that were common or unique to gene lists. However, when multiple gene lists were analysed, the identification of consistent underlying pathways or networks was more critical, because previous studies have reported that overlap between OMICs datasets was more readily apparent at the level of pathways or protein complexes [46,47]. Thus, the PLS1+ significant where probes exceed the background in at least 20% of the samples to perform the second set of enrichment analyses. ...
Aims: Generalized epilepsy is thought to involve distributed brain networks. However, the molecular and cellular factors that render different brain regions more vulnerable to epileptogenesis remain largely unknown. We aimed to investigate epilepsy-related morphometric similarity network (MSN) abnormalities at the macroscale level and their relationships with microscale gene expressions at the microscale level. Methods: We compared the MSN of genetic generalized epilepsy with generalized tonic-clonic seizures patients (GGE-GTCS, n = 101) to demographically-matched healthy controls (HC, n = 150). Cortical MSNs were estimated by combining seven morphometric features derived from structural magnetic resonance imaging for each individual. Regional gene expression profiles were derived from brain-wide microarray measurements provided by the Allen Human Brain Atlas. Results: GGE-GTCS patients exhibited decreased regional MSNs in primary motor, prefrontal, and temporal regions, and increases in occipital, insular, and posterior cingulate cortices, when compared to the HC. These case-control neuroimaging differences were validated using split-half analyses and were not affected by medication or drug response effects. When assessing associations with gene expression, genes associated with GGE-GTCS-related MSN differences were enriched in several biological processes, including "synapse organization", "neurotransmitter transport" pathways, and excitatory/inhibitory neuronal cell types. Collectively, the GGE-GTCS-related cortical vulnerabilities were associated with chromosomes 4, 5, 11, and 16, and were dispersed bottom-up at the cellular, pathway, and disease levels, which contributed to epileptogenesis, suggesting diverse neurobiologically relevant enrichments in GGE-GTCS. Conclusions: By bridging the gaps between transcriptional signatures and in vivo neuroimaging, we highlighted the importance of using MSN abnormalities of the human brain in GGE-GTCS patients to investigate disease-relevant genes and biological processes.
... In the natural environment mutations may arise spontaneously whereas under laboratory conditions, artificially induced genetic perturbations can be studied such as mutational screens, gene editing, and RNA interference (RNAi) [8]. This has provided a thorough understanding of molecular genetic pathways including the effect of genetic perturbation at large scale with multiple factors associated with a range of phenotypes in model organisms like yeast, flies, nematodes, and mice [9]. ...
... In parallel, there have been efforts to develop and optimize high-throughput analytic technologies such as RNAi [31][32][33] and CRISPR-based genetic perturbation screening [34][35][36][37][38][39][40][41][42][43] in order to systematically identify genetic vulnerabilities, with a prominent focus on cancer. In particular, CRISPR-screens have become an important method by which this is accomplished [32,33,41,42,[44][45][46][47][48][49][50]. ...
Genetic screens are powerful tools for both resolving biological function and identifying potential therapeutic targets, but require physiologically accurate systems to glean biologically useful information. Here, we enable genetic screens in physiologically relevant ex vivo cancer tissue models by integrating CRISPR-Cas-based genome engineering and biofabrication technologies. We first present a novel method for generating perfusable tissue constructs, and validate its functionality by using it to generate three-dimensional perfusable dense cultures of cancer cell lines and sustain otherwise ex vivo unculturable patient-derived xenografts. Using this system we enable large-scale CRISPR screens in perfused tissue cultures, as well as emulate a novel point-of-care diagnostics scenario of a clinically actionable CRISPR knockout (CRISPRko) screen of genes with FDA-approved drug treatments in ex vivo PDX cell cultures. Our results reveal differences across in vitro and in vivo cancer model systems, and highlight the utility of programmable tissue engineered models for screening therapeutically relevant cancer vulnerabilities.
... Even though RNAi screens are powerful techniques, they may result in a high rate of off-target effects. Therefore, RNAi screens have frequently been replaced by other techniques based on the CRISPR system in order to result in higher specificity and efficacy [98][99][100][101]. ...
Full-text available
Genome engineering makes the precise manipulation of DNA sequences possible in a cell. Therefore, it is essential for understanding gene function. Meganucleases were the start of genome engineering, and it continued with the discovery of Zinc finger nucleases (ZFNs), followed by Transcription activator-like effector nucleases (TALENs). They can generate double-strand breaks at a desired target site in the genome, and therefore can be used to knock in mutations or knock out genes in the same way. Years later, genome engineering was transformed by the discovery of clustered regularly interspaced short palindromic repeats (CRISPR). Implementation of CRISPR systems involves recognition guided by RNA and the precise cleaving of DNA molecules. This property proves its utility in epigenetics and genome engineering. CRISPR has been and is being continuously successfully used to model mutations in leukemic cell lines and control gene expression. Furthermore, it is used to identify targets and discover drugs for immune therapies. The descriptive and functional genomics of leukemias is discussed in this study, with an emphasis on genome engineering methods. The CRISPR/Cas9 system’s challenges, viewpoints, limits, and solutions are also explored.
... From this perspective, cells and their respective states are defined by a combination of transcription factors (TFs) that interact with a set of target genes to produce a specific gene expression profile. Perturbation of cell state through genetic, epigenetic, or pharmacologic means may provide insight into the contribution of these GRNs to cell function (Liberali et al., 2015;Li et al., 2018;Musa et al., 2018;Caldera et al., 2019). Diseases or pharmacologic treatments represent perturbations of GRNs and investigation of these states at the single cell level may shed light on critical GRNs that underlie human disease or responses to treatment by potentially identifying key regulators in these GRNs (del Sol et al., 2010;Fiers et al., 2018;Caldera et al., 2019). ...
Full-text available
The endocochlear potential (EP) generated by the stria vascularis (SV) is necessary for hair cell mechanotransduction in the mammalian cochlea. We sought to create a model of EP dysfunction for the purposes of transcriptional analysis and treatment testing. By administering a single dose of cisplatin, a commonly prescribed cancer treatment drug with ototoxic side effects, to the adult mouse, we acutely disrupt EP generation. By combining these data with single cell RNA-sequencing findings, we identify transcriptional changes induced by cisplatin exposure, and by extension transcriptional changes accompanying EP reduction, in the major cell types of the SV. We use these data to identify gene regulatory networks unique to cisplatin treated SV, as well as the differentially expressed and druggable gene targets within those networks. Our results reconstruct transcriptional responses that occur in gene expression on the cellular level while identifying possible targets for interventions not only in cisplatin ototoxicity but also in EP dysfunction.
... But many aspects of aging biology cannot be studied in terms of cellular abundance, and the variety of biological processes involved calls for multidimensional readouts. Single-cell RNA sequencing (scRNAseq) can provide this additional layer of interpretability for the effects of perturbations, similar to multivariate imaging screens (Liberali et al., 2015) but in a more realistic in vivo setting. Indeed, bulk transcriptomics has already been used not only to identify transcriptomic signatures of aging, but also to identify small molecule interventions that revert transcriptomes to a "younger" state and extend organismal lifespan (Baumgart et al., 2016;Janssens et al., 2019). ...
Full-text available
Biological aging, and the diseases of aging, occur in a complex in vivo environment, driven by multiple interacting processes. A convergence of recently developed technologies has enabled in vivo pooled screening: direct administration of a library of different perturbations to a living animal, with a subsequent readout that distinguishes the identity of each perturbation and its effect on individual cells within the animal. Such screens hold promise for efficiently applying functional genomics to aging processes in the full richness of the in vivo setting. In this review, we describe the technologies behind in vivo pooled screening, including a range of options for delivery, perturbation and readout methods, and outline their potential application to aging and age-related disease. We then suggest how in vivo pooled screening, together with emerging innovations in each of its technological underpinnings, could be extended to shed light on key open questions in aging biology, including the mechanisms and limits of epigenetic reprogramming and identifying cellular mediators of systemic signals in aging.
Bioinformatics is a rapidly advancing interdisciplinary field in which computational methods are applied to the analysis, interpretation, and visualization of biological and/or medical data, most often at the cell or molecular level and usually in the form of large, multivariate datasets. PubMed citations for “bioinformatics” have risen sharply, increasing 24‐fold since the year 2000. Massively parallel sequencing and imaging technologies are now generating a data deluge—many trillions of numbers in all—that strain the capacity of our hardware, software, and personnel infrastructures for bioinformatics; hence, the computational aspect of a sequencing project is often more expensive and time‐consuming than the laboratory work. The field of bioinformatics includes “data scientists,” for whom useful innovative algorithms, software, and data structures are publishable ends in themselves, and “computational biologists,” for whom algorithms, software, and computers are just tools of the trade for answering biological or medical questions. Bioinformatic analysis can be hypothesis‐generating or hypothesis‐testing, and a large number of statistical and artificial intelligence algorithms, scripts, and software packages are now available for each type of analysis. In fact, the rapidly increasing number of options is almost paralyzing. Computational consortia formed around large‐scale, public, molecular profiling projects like The Cancer Genome Atlas project have served as incubators for rapid computational advance, particularly driving major innovations in single‐cell informatics, artificial intelligence, and computational immunology. One attractive working arrangement is a stable “dyad” in which the bioinformatician focuses primarily on database management, statistical analysis, and data visualization, whereas the biologist/clinical researcher focuses primarily on interpretation and biomedical application.
Full-text available
Background Cell morphology is a complex and integrative readout, and therefore, an attractive measurement for assessing the effects of genetic and chemical perturbations to cells. Microscopic images provide rich information on cell morphology; therefore, subjective morphological features are frequently extracted from digital images. However, measured datasets are fundamentally noisy; thus, estimation of the true values is an ultimate goal in quantitative morphological phenotyping. Ideal image analyses require precision, such as proper probability distribution analyses to detect subtle morphological changes, recall to minimize artifacts due to experimental error, and reproducibility to confirm the results. Results Here, we present UNIMO (UNImodal MOrphological data), a reliable pipeline for precise detection of subtle morphological changes by assigning unimodal probability distributions to morphological features of the budding yeast cells. By defining the data type, followed by validation using the model selection method, examination of 33 probability distributions revealed nine best-fitting probability distributions. The modality of the distribution was then clarified for each morphological feature using a probabilistic mixture model. Using a reliable and detailed set of experimental log data of wild-type morphological replicates, we considered the effects of confounding factors. As a result, most of the yeast morphological parameters exhibited unimodal distributions that can be used as basic tools for powerful downstream parametric analyses. The power of the proposed pipeline was confirmed by reanalyzing morphological changes in non-essential yeast mutants and detecting 1284 more mutants with morphological defects compared with a conventional approach (Box–Cox transformation). Furthermore, the combined use of canonical correlation analysis permitted global views on the cellular network as well as new insights into possible gene functions. Conclusions Based on statistical principles, we showed that UNIMO offers better predictions of the true values of morphological measurements. We also demonstrated how these concepts can provide biologically important information. This study draws attention to the necessity of employing a proper approach to do more with less.
Full-text available
Image-based phenotypic screening relies on the extraction of multivariate information from cells cultured under a large variety of conditions. Technical advances in high-throughput microscopy enable screening in increasingly complex and biologically relevant model systems. To this end, organoids hold great potential for high-content screening because they recapitulate many aspects of parent tissues and can be derived from patient material. However, screening is substantially more difficult in organoids than in classical cell lines from both technical and analytical standpoints. In this review, we present an overview of studies employing organoids for screening applications. We discuss the promises and challenges of small-molecule treatments in organoids and give practical advice on designing, running, and analyzing high-content organoid-based phenotypic screens. Advances in culture techniques and three-dimensional image analysis algorithms are needed for robust chemical screening of organoids, tiny three-dimensional versions of functioning organs that are derived from stem cells. Organoids can provide useful insights into potential drug targets, cellular behavior, and disease progression, but their complexity provides unique challenges for image-based screening and analyses. Prisca Liberali at the Friedrich Miescher Institute for Biomedical Research in Basel, Switzerland, and co-workers reviewed the current status of image-based chemical screening for organoid-based biological and drug research. Careful assay design and process optimization are required, and the volume of generated data can place strain on computational infrastructure. The team highlights the value of multiplexed imaging to improve efficiency, and suggests that neural networks and deep learning may help tackle the complexities of three-dimensional image segmentation analyses.
Full-text available
DAVID bioinformatics resources consists of an integrated biological knowledgebase and analytic tools aimed at systematically extracting biological meaning from large gene/protein lists. This protocol explains how to use DAVID, a high-throughput and integrated data-mining environment, to analyze gene lists derived from high-throughput genomic experiments. The procedure first requires uploading a gene list containing any number of common gene identifiers followed by analysis using one or more text and pathway-mining tools such as gene functional classification, functional annotation chart or clustering and functional annotation table. By following this protocol, investigators are able to gain an in-depth understanding of the biological themes in lists of genes that are enriched in genome-scale studies.
Bacteria form architecturally complex communities known as biofilms in which cells are held together by an extracellular matrix. Biofilms harbor multiple cell types, and it has been proposed that within biofilms individual cells follow different developmental pathways, resulting in heterogeneous populations. Here we demonstrate cellular differentiation within biofilms of the spore-forming bacterium Bacillus subtilis, and present evidence that formation of the biofilm governs differentiation. We show that motile, matrix-producing, and sporulating cells localize to distinct regions within the biofilm, and that the localization and percentage of each cell type is dynamic throughout development of the community. Importantly, mutants that do not produce extracellular matrix form unstructured biofilms that are deficient in sporulation. We propose that sporulation is a culminating feature of biofilm formation, and that spore formation is coupled to the formation of an architecturally complex community of cells.
The functions of many open reading frames (ORFs) identified in genome-sequencing projects are unknown. New, whole-genome approaches are required to systematically determine their function. A total of 6925 Saccharomyces cerevisiae strains were constructed, by a high-throughput strategy, each with a precise deletion of one of 2026 ORFs (more than one-third of the ORFs in the genome). Of the deleted ORFs, 17 percent were essential for viability in rich medium. The phenotypes of more than 500 deletion strains were assayed in parallel. Of the deletion strains, 40 percent showed quantitative growth defects in either rich or minimal medium.
While the catalog of mammalian transcripts and their expression levels in different cell types and disease states is rapidly expanding, our understanding of transcript function lags behind. We present a robust technology enabling systematic investigation of the cellular consequences of repressing or inducing individual transcripts. We identify rules for specific targeting of transcriptional repressors (CRISPRi), typically achieving 90%-99% knockdown with minimal off-target effects, and activators (CRISPRa) to endogenous genes via endonuclease-deficient Cas9. Together they enable modulation of gene expression over a ∼1,000-fold range. Using these rules, we construct genome-scale CRISPRi and CRISPRa libraries, each of which we validate with two pooled screens. Growth-based screens identify essential genes, tumor suppressors, and regulators of differentiation. Screens for sensitivity to a cholera-diphtheria toxin provide broad insights into the mechanisms of pathogen entry, retrotranslocation and toxicity. Our results establish CRISPRi and CRISPRa as powerful tools that provide rich and complementary information for mapping complex pathways.
Endocytosis is critical for cellular physiology and thus is highly regulated. To identify regulatory interactions controlling the endocytic membrane system, we conducted 13 RNAi screens on multiple endocytic activities and their downstream organelles. Combined with image analysis of thousands of single cells per perturbation and their cell-to-cell variability, this created a high-quality and cross-comparable quantitative data set. Unbiased analysis revealed emergent properties of the endocytic membrane system and how its complexity evolved and distinct programs of regulatory control that coregulate specific subsets of endocytic uptake routes and organelle abundances. We show that these subset effects allow the mapping of functional regulatory interactions and their interaction motifs between kinases, membrane-trafficking machinery, and the cytoskeleton at a large scale, some of which we further characterize. Our work presents a powerful approach to identify regulatory interactions in complex cellular systems from parallel single-gene or double-gene perturbation screens in human cells and yeast.