Co Expresion

Briefings in Bioinformatics, 19(4), 2018, 575–592
doi: 10.1093/bib/bbw139
Advance Access Publication Date: 10 January 2017
Paper
Gene co-expression analysis for functional
Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

classification and gene–disease predictions
Sipko van Dam, Urmo Vo ~ sa, Adriaan van der Graaf, Lude Franke, and
~ o Pedro de Magalha
Joa ~ es
Corresponding authors: Sipko van Dam, Systems Genetics, Department of Genetics, UMCG HPC CB50, P.O. Box: 30001, 9700 RB Groningen,
The Netherlands, Tel.: þ31 50 361 72 29; Fax: þ31 50 361 72 31; E-mail: [email protected]; Joa ~ es, Institute of Ageing and Chronic
Disease, University of Liverpool, William Duncan Building, Room 281, 6 West Derby Street, Liverpool L7 8TX, United Kingdom, Tel.: þ44 151 7954517;
Fax: þ44 151 795 8420; E-mail: [email protected]
Abstract
Gene co-expression networks can be used to associate genes of unknown function with biological processes, to prioritize
candidate disease genes or to discern transcriptional regulatory programmes. With recent advances in transcriptomics and
next-generation sequencing, co-expression networks constructed from RNA sequencing data also enable the inference of
functions and disease associations for non-coding genes and splice variants. Although gene co-expression networks typic-
ally do not provide information about causality, emerging methods for differential co-expression analysis are enabling the
identification of regulatory genes underlying various phenotypes. Here, we introduce and guide researchers through a (dif-
ferential) co-expression analysis. We provide an overview of methods and tools used to create and analyse co-expression
networks constructed from gene expression data, and we explain how these can be used to identify genes with a regulatory
role in disease. Furthermore, we discuss the integration of other data types with co-expression networks and offer future
perspectives of co-expression analysis.
Key words: transcriptomics; functional genomics; disease gene prediction; next-generation sequencing; network analysis
Introduction sequencing (RNA-seq), and their respective data-analysis meth-

ods, the functional status of a gene can now be identified from a
A key objective in biological research is to systematically identify systematic perspective [2, 3]. One method to infer gene function
all molecules within a living cell and how they interact. However, and gene–disease associations from genome-wide gene expres-
the functions of many genes are still not understood, a situation sion is co-expression network analysis (Figure 1), an approach
that has only become more complex with the recent identifica- that constructs networks of genes with a tendency to co-activate
tion of many novel non-coding genes [1]. With the development across a group of samples and subsequently interrogates and
of high-throughput technologies including microarrays and RNA analyses this network.
Sipko van Dam is a researcher at the Department of Genetics, UMC Groningen. He carried out his doctoral work at the University of Liverpool creating and
analysing a co-expression network constructed from public RNA-seq data.
Urmo Vo ~ sa is researcher in the Department of Genetics, UMC Groningen. His main interests lie in the genetics of gene expression and integration of data
from different layers of genomic complexity to untangle the causes of complex diseases.
Adriaan van der Graaf is a master’s student of Molecular Biology and Biotechnology at the Department of Genetics, UMC Groningen, focusing on novel
statistical techniques in the analysis of expression data.
Lude Franke is an associate professor at the Department of Genetics at the University Medical Centre Groningen. He is a statistical geneticist, working on
analysing data on the genetics of complex and autoimmune diseases (e.g. celiac disease).
Joa ~ es is a reader at the University of Liverpool where he leads the Integrative Genomics of Ageing Group (http://pcwww.liv.ac.uk/aging/).
The group’s research integrates different strategies but its focal point is developing and applying experimental and computational methods that help
bridge the gap between genotype and phenotype, and help decipher the human genome and how it regulates complex processes like ageing.
Submitted: 12 September 2016; Received (in revised form): 1 December 2016
C The Author 2017. Published by Oxford University Press.
V
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/),
which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
575
576 | van Dam et al.

Figure 1. Example of a co-expression network analysis. First, pairwise correlation is determined for each possible gene pair in the expression data. These pairwise correl-
ations can then be represented as a network. Modules within these networks are defined using clustering analysis. The network and modules can be interrogated to iden-
tify regulators, functional enrichment and hub genes. Differential co-expression analysis can be used to identify modules that behave differently under different
conditions. Potential disease genes can be identified using a guilt-by-association (GBA) approach that highlights genes that are co-expressed with multiple disease genes.
Gene co-expression networks can be used for various pur- expression methods are sensitive to noise [21], they are becoming
poses, including candidate disease gene prioritization, func- more effective with the increase in RNA-seq data quantity and
tional gene annotation (Figure 1) and the identification of quality. RNA-seq further permits co-expression analysis to focus
regulatory genes. However, co-expression networks are effect- on splice variants and non-coding RNAs.
ively only able to identify correlations; they indicate which genes In this review, we provide an introduction and overview of
are active simultaneously, which often indicates they are active what constitutes a co-expression network, followed by a guide
in the same biological processes, but do not normally confer in- of the different steps in co-expression analysis using RNA-seq
formation about causality or distinguish between regulatory and data. We then describe commonly used and newly emerging
regulated genes. An increasingly used method that goes beyond methods and tools for co-expression analysis, with a focus on
traditional co-expression networks is differential co-expression differential co-expression analysis to identify regulatory genes
analysis [4–7]. This approach identifies genes with varying co- that underlie disease. We conclude with a discussion of the in-
expression partners under different conditions, such as disease tegration of co-expression networks with other types of data, to
states [4, 8–10], tissue types [11] and developmental stages [12], e.g. infer regulatory processes, and with future prospects and
because these genes are more likely to be regulators that under- remaining challenges in the field.
lie phenotypic differences. The regulatory roles of such genes
can be further investigated by integrating data types such as pro-
Co-expression networks
tein–protein interactions, methylome data, interactions between
transcription factors (TFs) and their targets, and with sequence A co-expression network identifies which genes have a tendency
motif analysis of co-expressed genes [13–15]. This aids in the to show a coordinated expression pattern across a group of sam-
identification of regulatory elements such as TFs, expression ples. This co-expression network can be represented as a gene–
quantitative trait loci (eQTLs) and methylation patterns that af- gene similarity matrix, which can be used in downstream ana-
fect the expression and composition of co-expression modules. lyses (Figure 1). Canonical co-expression network construction
Gene expression and regulation can be highly tissue-specific, and analyses can be described with the following three steps.
and most disease-related genes have tissue-specific expression In the first step, individual relationships between genes are
abnormalities [16, 17]. The increased availability of expression data defined based on correlation measures or mutual information
for multiple tissues has allowed for differential co-expression ana- [22–24] between each pair of genes. These relationships describe
lysis, which can identify both tissue-specific signatures and shared the similarity between expression patterns of the gene pair
co-expression signatures [11]. These tissue-specific signatures can across all the samples. Different measures of correlation have
be disrupted in tissue-specific diseases and would not be detected been used to construct networks, including Pearson’s or
in analyses aggregating multiple tissues. Even when no sample Spearman’s correlations [25, 26]. Alternatively, least absolute
classification is available, subpopulation-specific modules can be error regression [27] or a Bayesian approach [28] can be used to
resolved, an approach that has been particularly successful in clas- construct a co-expression network. The latter two have the
sifying different cancer subtypes to provide prognostic markers added benefit that they can be used to identify causal links and
[18–20]. Differential co-expression analysis is also useful for analy- have been explained elsewhere [29]. For a discussion of other
sing data sets in which the subpopulations are unknown, e.g. types of similarity measures, we refer to [30]. Many of these
large-scale single-cell RNA-seq data [5, 12]. While differential co- similarity metrics can also be used to construct protein–protein
Gene co-expression analysis | 577
interaction networks, which were compared using cancer data gene pairs is binary, i.e. either 0 or 1, and genes are either con-
in [31]. nected or unconnected. An un-weighted network can be created
In the second step, co-expression associations are used to from a weighted network by, for example, considering all genes
construct a network where each node represents a gene and with a correlation above a certain threshold to be connected
each edge represents the presence and the strength of the co- and all others unconnected. We focus on weighted networks in
expression relationship (Figure 1) [32]. this review because (to date) they have produced more robust
In the third step, modules (groups of co-expressed genes) are results than un-weighted networks [40].
identified using one of several available clustering techniques.
Clustering in co-expression analyses is used to group genes Microarrays versus RNA-seq data
with similar expression patterns across multiple samples to
Co-expression networks can be constructed from gene expres-

produce groups of co-expressed genes rather than only pairs.
The clustering method needs to be chosen with consideration sion data obtained from microarray or RNA-seq technology.
because it can greatly influence the outcome and meaning of One of the major benefits of RNA-seq is that it quantifies the ex-
the analysis. Many clustering methods are available, including pression of the over 70 000 non-coding RNAs not usually meas-
k-means clustering and hierarchal clustering, and are discussed ured with microarrays [1], including recently annotated
in detail in [33]. Modules can subsequently be interpreted by lincRNAs, many of which are thought to have regulatory roles
functional enrichment analysis, a method to identify and rank [41] and to play a role in disease [42, 43]. Therefore, to gain a bet-
overrepresented functional categories in a list of genes [34–36]. ter understanding of the regulatory mechanisms driving biolo-
In co-expression analysis, it is important to consider the gical processes, non-coding RNAs need to be considered in
heterogeneity of the samples. Tissue-specific or condition- analyses.
specific co-expression modules may not be detectable in a co- RNA-seq also has other benefits [35]. It increases accuracy
expression network constructed from multiple tissues or for low-abundance transcripts [44], has a higher resolution for
conditions because the correlation signal of the tissue/ identifying tissue-specific expression and distinguishes expres-
condition-specific modules is diluted by a lack of correlation in sion profiles of closely related paralogues better than
other tissues/conditions. However, limiting co-expression ana- microarray-derived profiles [45]. RNA-seq can also distinguish
lysis to a specific tissue or condition also reduces sample size, between the expression of different splice variants [46, 47],
thereby also decreasing the statistical power to detect shared which can have distinct interaction partners [48] and biological
co-expression modules. Therefore, methods that do not distin- functions [49]. Co-expression analysis on RNA-seq data can as-
guish between tissues or conditions should be used for identifi- sign putative roles to these splice variants and lincRNAs [2], and
cation of common co-expression modules, while differential identify diseases in which they might play a part [2]. A limita-
co-expression comparing different conditions or tissues will be tion of co-expression analysis on the splice variant level is the
better for identifying modules unique to a specific condition or introduction of biases because it is difficult to determine which
tissue. splice variant is expressed if multiple splice variants share the
same expressed exon.
As an example of RNA-seq’s utility with isoform- and exon-
Types of co-expression networks
specific expression level measurements, exon-level expression
Signed and unsigned co-expression networks was used to construct a co-splicing network [50, 51]. In a gene
In a correlation-based co-expression network, correlation meas- co-expression network, expression of different transcripts origi-
ures have values between 1 (perfect negative correlation) and nating from the same gene is usually aggregated, which can
1 (perfect positive correlation). In an unsigned network, the ab- lead to biased co-expression signals [50]. In a co-splicing net-
solute correlation values are used, which means that two nega- work, this issue is resolved by considering the exon-expression-
tively correlated genes will be considered as co-expressed. This level distributions within a gene when calculating gene co-
causes negatively correlated genes to group together. Because expression correlation. In biological terms this means that the
those genes are likely to be also positively co-expressed with a expression of two genes is only considered to be correlated if
completely different set of genes, these genes also group into their different splice variants show co-ordinated expression. If
the same module and disrupt the structure of the network. A this is not the case, they are not considered to be co-expressed
signed network solves this problem by scaling the correlation even if the overall expression levels of the genes are correlated.
values between 0 and 1 so that values <0.5 indicate negative This approach has identified novel functional modules, which
correlation and values >0.5 indicate positive correlation. A would not be detected using traditional co-expression networks
signed method creates networks where biologically meaningful [51]. Additionally, genes that contain multiple exons and tran-
modules (such as those representing a specific biological pro- scripts acquired more relevant positions in the network using
cess) are better separated [37]. Thus, a scaled value close to 0 in- this method [50], a reassuring result given that splice variants
dicates negative correlation, a feature which may be can have different functions and are thus likely co-expressed
particularly interesting when microRNAs (miRNAs) are incorpo- with functionally distinct partners, which co-splicing networks
rated into the network, as these are known to exert their func- account for.
tion mainly through down-regulation of other genes [38]. This A different approach is to determine the expression of differ-
also holds true for some long intergenic non-coding RNAs ent isoforms originating from the same gene based on the dis-
(lincRNAs) [39]. tributions of reads mapping to its various exons. This method is
used by SpliceNet, which effectively divides the reads mapping
Weighted and un-weighted co-expression networks to an exon shared with two isoforms proportionally to the total
In a weighted network, all genes are connected to each other, expression of each of the two whole isoforms [52]. This means
and these connections have continuous weight values between that if two isoforms, isoform A and isoform B, share only one
0 and 1 that indicate the strength of co-regulation between the exon X (to which a number of reads map), but there are no reads
genes. In an un-weighted network, the interaction between mapping to the other exons of isoform A, whereas some reads
map to the exons of isoform B, all reads mapping to exon X are co-expression studies have used a cut-off of 10 million reads
then assigned to isoform B, resulting in isoform A being con- per sample [2, 21, 60]. Co-expression networks constructed
sidered as not expressed at all. Although this elegant solution using this cut-off have been suggested to have a similar quality
was validated using simulations, no experimental validation to microarray-based co-expression networks if constructed
was conducted. from the same number of samples [21], but decreasing in quality
The most common way of constructing RNA-seq-based co- with fewer reads. The percentage of mapped reads is another
expression networks is to merge all overlapping gene isoforms frequently considered cut-off in which samples with <70% or
in the RNA-seq data analysis and then construct the network at 80% of the reads mapping to the genome are removed. Giorgi
the gene level. This approach, however, loses information about et al. demonstrated, using 65 Arabidopsis thaliana samples with
different transcripts encoded by the same gene. Alternatively, 12 million reads but applying only a 30% mapping cut-off

transcript-based co-expression networks can be constructed. threshold, that the resulting RNA-seq-based co-expression net-
The drawback of these networks is their dramatic increase in work had a lower similarity to biological networks than micro-
size owing to the many gene isoforms and non-coding RNAs. As array networks [61]. Cut-off thresholds may vary per species,
co-expression networks are square matrixes, the size of the net- based on, among other factors, the quality of the genome anno-
work increases quadratically (n2) with the number of genes tation. As more and higher quality data become available,
included. As there are 200 000 annotated transcripts in the higher cut-off thresholds may be preferable.
human genome (according to Ensembl GRCh38.p5 (human) an- To ensure that a network is robust, bootstrapping can be
notation [53]) and only 20 000 protein-coding genes, the result- used [62]. This is the repetitive construction of networks by
ing network increases 100-fold in size, greatly increasing the using random sets of samples (one sample can be part of mul-
computational resources needed for the analysis. One solution tiple subsets) from the data, which are subsequently used to as-
to this problem is to build co-expression network blocks from sess the reproducibility of the network created from the entire
subsets of the data and combine these blocks at a later point in data set. Randomizing the data set (e.g. by randomly reassign-
the analysis [54]. We recommend users to be cautious with ing expression values to their gene/transcript identifiers and re-
block-wise clustering, however, as it may influence the results constructing the network) can also help identify correlations
of subsequent module detection analyses, and it is unclear how that occur stochastically because of specific biases rather than
well these perform when large numbers of blocks are used. as a result of biologically relevant interactions [2].
RNA-seq data for co-expression networks Clustering and network analysis

RNA-seq analysis entails multiple steps that include obtaining Identifying modules
expression estimates from the sequenced reads, data normal-
ization and quality control. Different tools and methods to ob- Clustering is used to group genes that have a similar expression
tain reliable expression counts from RNA-seq data were pattern in multiple samples. The resulting modules often represent
recently reviewed in [55], and these will not be reviewed here. biological processes [63, 64] and can be phenotype specific [65].
In our experience, different normalization methods intro- The most widely used clustering package for co-expression
duce different biases in co-expression analysis, usually towards analysis is Weighted Gene Correlation Network Analysis
positive correlation. New methods are continuously being cre- (WGCNA) [40]. This easy-to-use tool constructs co-expression
ated to tackle these normalization issues. The recently pub- modules using hierarchical clustering on a correlation network
lished method extracting patterns and identifying co-expressed created from expression data [54]. Hierarchical clustering itera-
genes (EPIG) from RNA-Seq data (EPIG-seq), for example, is de- tively divides each cluster into sub-clusters to create a tree with
signed to calculate gene correlation across RNA-seq samples, branches representing co-expression modules. Modules are
being unaffected by read-depth differences between samples then defined by cutting the branches at a certain height
and the large abundance of 0 values present in RNA-seq- (Figure 1).
derived expression matrices [56]. Biases originating from the WGCNA was the first co-expression tool to be applied to
large abundance of 0 values are even more pronounced in RNA-seq data; it has effectively identified biologically relevant
single-cell experiments because of low RNA quantities per cell. associations between phenotypes and modules [19, 66, 67], per-
Specific tools have been created for analysis of single-cell RNA- forming similarly to microarray-based analyses. An RNA-seq-
seq data and are reviewed in [57]. Although some studies com- based co-expression study on normal and failing murine hearts
paring different normalization methods for RNA-seq data are found that many lincRNAs are present in clusters correlating
available [58], more comprehensive comparison studies incor- with the failing murine heart phenotype, suggesting a possible
porating newer methods are needed. role of these non-coding RNAs in this disease [67]. Co-
expression analysis of RNA-seq data of lean and obese porcines
Minimum read depth and sample size required for identified obesity-related modules [66], and a link was found be-
co-expression analyses tween obesity, the immune system and bone remodelling, with
To create co-expression networks from RNA-seq data, a 20-sam- the study identifying CCR1, MSR1 and SPI1 as possible regulators
ple minimum has been suggested [21, 54], and increased sample in these processes. WGCNA was also used to identify biologic-
sizes produce networks with a higher functional connectivity ally relevant associations from single-cell RNA-seq data.
[21, 59]. Not surprisingly, higher quality data tend to result in Regulatory mechanisms and genes underlying pre-
more accurate co-expression networks [21, 59]. It is therefore es- implantation processes conserved between humans and mice
sential to set cut-off thresholds for data quality control. A were identified by using preservation detection defined by
higher total read depth for RNA-seq samples increases the ac- WGCNA [12], a feature that was later added to this package [68].
curacy of the expression measurements, especially for genes Co-expression modules were identified for different develop-
with low expression [21, 59]. For RNA-seq data, sequencing mental stages of human and mice separately. The modules
depth cut-off thresholds are usually selected arbitrarily. Several identified for each stage were then compared between humans
and mice to reveal a strong overlap between co-expression Guilt by association

modules in oocyte formation in mice and oocyte and single-cell
A widely used approach to attach biological meaning to mod-
stage co-expression modules in humans. This suggests that
ules is to determine functional enrichment among the genes
humans and mice share core transcriptional programmes in within a module using e.g. the tools described in Table 1.
early development, but diverge at a later stage [12]. Assuming that co-expressed genes are functionally related, en-
riched functions can be assigned to poorly annotated genes
Identifying hub genes within the same co-expression module, an approach commonly
Co-expression modules identified by clustering are often large, referred to as ‘guilt by association’ (GBA) [121]. GBA approaches
and so, it is important to identify which gene(s) in each module are also widely used to identify new potential disease genes if a
substantial proportion of the genes within a module are associ-

best explains its behaviour. A widely used approach is to iden-
ated with a particular disease [26, 121–126] (Figure 1).
tify highly connected genes in a co-expression network (hub
When using a GBA approach it is important to remember
genes). Hubs are frequently more relevant to the functionality
that not every gene in a module necessarily correlates with a
of networks than other nodes [69]. This is also the case in biolo-
function or disease association for which it is enriched. Because
gical networks [32], although mathematical derivations show
co-expression modules often consist of a large number of genes,
that this is only the case for intra-modular hub genes (as
any overrepresentation of a functional process or group of
opposed to inter-modular hub genes [64, 65]). Intra-modular
disease-associated genes quickly becomes statistically signifi-
hubs are central to specific modules in the network, while inter-
cant, as often indicated by deceivingly low p-values.
modular hubs are central to the entire network (Figure 2). To
Misinterpretation of these low p-values may lead to the incor-
identify hub genes, centrality measures, mainly ‘betweenness rect conclusion that all genes in a module play an important
centrality’, are often used. Genes with high betweenness cen- part in a particular process or disease. In reality, the fraction of
trality are important as shortest-path connectors through a net- genes in a module that relate to its main biological function is
work [70]. Connectivity is often used to measure network often <20% [127], and module-trait correlations can be relatively
robustness and indicates how many genes need to be removed low (correlation < 0.5) even when statistically significant [128].
from the network before the remaining genes are disconnected.
Identifying hub genes in co-expression networks has led to the Regulatory network construction
identification of several genes essential in cancer [71, 72], type 2 Although there is ample evidence that co-expression analysis
diabetes [73], chronic fatigue [74], other diseases [75, 76] and tis- can help identify genes that play an important role in disease
sue regeneration [77]. and biological functions, it remains difficult to infer causality
As there are usually multiple hubs or differentially con- from co-expression networks. Tools such as ARACNE [23] and
nected genes in a module, it is not always clear which is the GENIE3 [113] attempt to construct regulatory networks from co-
most important gene underlying a phenotype. Nor is there a expression networks. ARACNE removes indirect connections
guarantee that any of the hub genes is causal to a phenotype between genes (i.e. partners of a gene that have a stronger cor-
[78]. One study using myocardial data from 1617 samples found relation with each other than with the gene itself), leaving only
that known foetal gene markers upregulated in gene networks those connections that are expected to be regulatory. GENIE3 in-
common to developing and diseased myocardium were not hub corporates TF information to construct a regulatory network by
genes [79]. Another co-expression study in Salmonella found that determining the TF expression pattern that best explains the
hub genes are dispensable for growth, stress adaptation and expression of each of their target genes. A limitation of GENIE3
virulence, suggesting that hub genes are not necessarily essen- is that TF information is required for it to perform better than
tial [80]. random chance [113]. The performance of these methods has
been compared with gold standards defined by regulatory inter-
actions experimentally validated in >150 studies. The compari-
son suggests that methods attempting to derive regulatory
networks from co-expression networks alone can only reliably
distinguish between true- and false-positive regulatory inter-
actions if perturbation experiment data are used for network
construction [129]. A comparison between these tools and
others, including WGCNA, showed that WGCNA and ARACNE
perform best at defining the network structure of Escherichia coli
[130], for which a well-defined regulatory network was used as a
gold standard [131].
Differential co-expression analysis

Differential co-expression analysis can identify biologically im-
portant differential co-expression modules that would not be
detected using regular co-expression or differential expression
analyses. Genes that are differentially co-expressed between
Figure 2. Hypothetical network explaining inter- and intra-modular hubs and different sample groups are more likely to be regulators, and
network centrality. The inter-modular hub has a high network centrality, as it is
are therefore likely to explain differences between phenotypes
required for the largest number of shortest paths between all possible node
pairs. The red line indicates an example of a shortest path through the network
[4, 8–10]. Differential co-expression analysis has been used to
between a pair of nodes. Intra-modular hubs (marked with orange) are central identify genes underlying differences between healthy and dis-
to individual modules and usually have high biological relevance. ease samples [4, 8–10] or between different tissues [11], cell
Table 1. Methods and tools for RNA-seq-based co-expression network analysis
Tool/method Description, strengths (þ) and limitations ()
Quality control
FastQC [81] A tool that uses .fastq, .bam or .sam files to identify and highlight potential issues in the
http://www.bioinformatics.babraham. data, such as low base quality scores, low sequence quality and GC content biases.
ac.uk/projects/fastqc/ þ Can be used either with or without user interface.
Uses only the first 200 000 sequences in the file.
RSeQC [82] þ A tool with a wider range of quality control measures than FastQC.
http://rseqc.sourceforge.net/ þ Can also be used on mapped data to obtain information on metrics such as the preva-

lence of splicing events.
QoRTs [83] þ This is a similar tool to RSeQC but incorporates more quality control metrics.
http://hartleys.github.io/QoRTs/
Read Mappers
Bowtie/Tophat/Tophat2 [84] The first widely used mapping tool.
https://ccb.jhu.edu/software/tophat/ þ Detects splice variants.
index.shtml Currently much slower than most other mappers and requires a relatively large amount
of memory.
STAR [85] A widely used tool to align reads to a genome.
https://code.google.com/p/rna-star/ þ Maps 50 times faster than Tophat and Tophat2.
þ Commonly used tool to detect novel splice variants.
Uses a large amount of memory (>20 GB for mapping to the human genome).
HISAT [86] A widely used tool to align reads to a genome at a faster rate than STAR with comparable
http://www.ccb.jhu.edu/software/hisat/ accuracy.
index.shtml þ HISAT2 is expected to be the core of the next version of Tophat (Tophat3).
þ Detects novel splice variants.
þ The newer HISAT2 version aligns to genotype variants, likely achieving higher accuracy.
þ Uses less memory than STAR (<8 GB for mapping to the human genome using default
settings).
BWA [87] A commonly used aligner for species in which splicing does not occur.
Does not detect splice variants.
Kallisto [88] A tool that uses a pseudoalignment strategy to assign expression values to transcripts/
https://pachterlab.github.io/kallisto/ genes to achieve optimal speed.
about.html Comparable accuracy to other tools using real alignment strategies.
Reports reads/expression per gene instead of read alignment coordinates (which are com-
monly used to acquire the expression per gene).
þ Uses little memory and can be run on a regular desktop computer.
Does not identify novel splice variants
Salmon [89] Another pseudoalignment tool. Performance comparable with Kallisto.
http://combine-lab.github.io/salmon/ Reports reads/expression per gene instead of read alignment coordinates (which are com-
monly used to acquire the expression per gene).
Does not identify novel splice variants.
Read counting tools
HTseq [90] A tool that assigns expression values to genes based on reads that have been aligned
http://www-huber.embl.de/HTSeq/doc/ with, e.g. STAR or HISAT.
overview.html þ Well documented and supported.
FeatureCounts [91] þ A tool that is similar to HTseq but much faster. Results are slightly different owing to
http://bioinf.wehi.edu.au/ slightly different expression assignment strategies.
featureCounts/
SpliceNet [52] A tool that divides the reads mapping to an exon shared with two isoforms proportionally
http://jjwanglab.org/SpliceNet/ to the total expression of each of the two whole isoforms.
þ Estimates expression more accurately when multiple genes/transcripts partly share the
same genome regions.
Normalization
FPKM/RPKM [92] Widely used normalization methods that correct for the total number of reads in a sam-
ple while accounting for gene length.
TMM has been suggested as a better alternative [58].
TPM [93] A method similar to FPKM, but normalizes the total expression to 1 million, i.e. the
summed expression of TPM-normalized samples is always 1 million.
TMM [94] Similar to FPKM/RPKM but puts expression measures on a common scale across different
samples.
RAIDA [95] A method that uses ratios between counts of genes in each sample for normalizations.
þ Avoids problems caused by differential transcript abundance between samples (resulting
from differential expression of highly abundant gene transcripts).
DEseq2 [96] A normalization method that adjusts the expression values of each gene in a sample by a
set factor. This factor is determined by taking the median gene expression in a sample
(continued)
Table 1. Continued
after dividing the expression of each gene by the geometric mean of the given gene
across all samples. This differs from the normalization implemented in the DEseq2 dif-
ferential expression analysis.
Implemented into the DEseq2 R package.
Correction for batch effects
Limma-removeBatchEffect [97] A method which uses linear models to correct for batch effects.
Svaseq [98] This method estimates biases based on genes that have no phenotypic expression effects,
https://github.com/jtleek/svaseq which are then used for correction of the data.

Specifically designed for RNA-seq data.
Combat [99] A method that is robust to outliers and also effective at batch effect correction in small
http://www.bu.edu/jlab/wp-assets/ sample sizes (<25).
ComBat/Abstract.html
Co-expression module detection
WGCNA [54] A tool that constructs a co-expression network using Pearson correlation (default) or a
https://labs.genetics.ucla.edu/horvath/ custom distance measure.
CoexpressionNetwork/Rpackages/ Uses hierarchical clustering and has various ‘tree cutting’ options to identify modules.
WGCNA/ þ Most widely used tool, well supported and documented.
DiffCoEx [100] A method that uses a similar approach to WGCNA to identify and group differentially co-
expressed genes instead of identifying co-expressed modules.
Identifies modules of genes that have the same different partners between different samples.
DICER [4] A method that identifies modules that correlate differently between sample groups, e.g.
modules that form one large interconnected module in one group compared with several
smaller modules in another group.
CoXpress [101] A tool that identifies co-expression modules in each sample group and tests whether the
http://coxpress.sourceforge.net/ genes within these modules are also co-expressed in other groups.
DINGO [102] DINGO is a more recent tool that groups genes based on how differently they behave in a
particular subset of samples (representing e.g. a particular condition) from the baseline
co-expression determined from all samples
GSCNA [103] A tool that tests whether a predefined defined gene set is differentially expressed between
two sample groups.
GSVD [104] A method that identifies ‘genelets’, which can be interpreted as modules representing
partial co-expression signals from multiple genes. These signals are then compared be-
tween two groups to identify genelets unique to samples and genelets that are shared be-
tween the two groups.
HO-GSVD [105] A tool similar to GSVD, but that can be used across multiple sample groups rather than
https://github.com/aanchan/hogsvd-py only two.
thon/blob/master/README.md
Biclustering [106] A group of methods that identify modules that are unique to a subpopulation of samples
without the need for prior grouping of samples.
Functional enrichment
DAVID [107] A widely used tool with an online web interface. Users supply a list of genes and select
https://david.ncifcrf.gov/ the annotation categories from various sources to identify enrichment.
PANTHER [108] A tool that uses a comprehensive protein library combined with human curated pathways
http://pantherdb.org/ and evolutionary ontology.
If a gene is not in the library, it is classified based on its protein sequence conservation
and by finding a related gene.
g:Profiler [109] A tool that performs enrichment analyses for gene ontologies, KEGG pathways, protein–
http://biit.cs.ut.ee/gprofiler/ protein interactions, TF and miRNA binding sites.
þ Also available as an R package.
ClusterProfiler [110] An R package for overrepresentation and gene set enrichment analyses for several cura-
https://github.com/GuangchuangYu/ ted gene sets.
clusterProfiler/blob/master/vignettes/ þ Allows users to compare the results of analyses performed on several gene sets.
clusterProfiler.Rmd
Enrichr [111] An intuitive web tool for performing gene overrepresentation analyses using a compre-
http://amp.pharm.mssm.edu/Enrichr/ hensive set of functional annotations.
ToppGene [36] An intuitive tool that determines enrichment of different categories such as GO terms,
https://toppgene.cchmc.org/ chromosomal locations and disease associations.
þ Also has other functions, such as candidate gene prioritization, based on network
structures.
Regulatory network inference
ARACNE [112] A tool that removes indirect connections between genes (i.e. partners of a gene that have
a stronger correlation with each other than with the gene itself), leaving only those con-
nections that are expected to be regulatory.
þ Creates directional networks.
(continued)
Table 1. Continued
Genie3 [113] A tool that incorporates TF information to construct a regulatory network by determining
the TF expression pattern that best explains the expression of each of their target genes.
þ Creates directional networks.
Requires TF information.
CoRegNet [114] A tool that identifies co-operative regulators of genes from different data types.
cMonkey [115] Calculates joint bicluster membership probability from different data types by identifying
groups of genes that group together in multiple data types.
Visualization

Cystoscape [116] A widely used tool for the visualization of networks.
http://www.cytoscape.org/ þ Has many plug-ins available for specific analyses.
BioLayout [117] Similar to Cytoscape but less widely used.
http://www.biolayout.org/ þ Can load and visualize much larger networks than Cytoscape.
Co-expression databasesa
COXPRESdb [60] A web resource incorporating 12 co-expression networks for different species created
http://coxpresdb.jp/ from 157 000 microarrays and 10 000 RNA-seq samples. Has a focus on protein-coding
RNAs.
GeneFriends [2] Human and mouse gene and transcript co-expression networks.
http://www.genefriends.org/ Networks constructed from 4000 RNA-seq samples each.
þ Includes a number of non-coding RNAs (10 000 for mouse and 25 000 for human).
GeneMANIA [118] Also includes physical and genetic interaction, co-localization, pathway and shared pro-
http://www.genemania.org/ tein domain information data sets.
þ Networks for nine species.
GENEVESTIGATOR [119] A database constructed using 145 000 samples.
https://genevestigator.com/gv/ þ Curated database.
þ Networks for 18 species.
þ Multiple data types.
GIANT [120] Tissue-specific interaction network database.
http://giant.princeton.edu/ Includes 987 Datasets encompassing 38 000 conditions describing 144 tissues types.
þ Integrates physical interaction, co-expression, miRNA binding motif and TF binding site
data.
This is a non-comprehensive list of available tools and methods.

a
These databases can be queried for a gene or multiple genes of interest to identify commonly co-expressed genes across the samples the database was created from.
types [5] or species [132, 133]. Below, we provide an overview of the analysis. It then prioritizes which genes in these modules
commonly used and newly emerging methods and tools, sepa- are likely to underlie the phenotype associated with the module
rated into two categories: (1) approaches that identify differen- by identifying either genes behaving similarly to the eigengene
tial co-expression between predefined sample groups (such as of the module or those genes that are intra-modular hub genes
conditions, time points or tissue types) and (2) approaches that (these tend to coincide). By design, DICER is tailored to identify
do not require prior knowledge about sample groups and use an module pairs that correlate differently between sample groups,
algorithm that identifies co-expression clusters in a priori un- e.g. modules that form one large interconnected module in one
known subpopulations of the samples. group compared with several smaller modules in another
(Figure 3D). DICER may be particularly useful for time series ex-
Differential co-expression analysis between periments in which co-expression changes are gradual, e.g. cell
cycle series experiments, where modules are specific to a par-
sample groups
ticular phase and co-expressed in transitions between phases.
Most differential co-expression analyses rely on differential DiffCoEx focuses on modules that are differentially co-
clustering; they identify clusters that contain different genes or expressed with the same sets of genes. The most extreme case
behave differently under changing conditions or phenotypes. of this behaviour is sets of genes that ‘hop’ from one set of cor-
The most frequently used programs for differential clustering related genes to another in a coordinated manner (Figure 3E). In
analysis, which have also been compared with others programs, this case, DiffCoEx would cluster ‘hopping’ genes in a similar
are WGCNA [54], DICER [4] and DiffCoEx [100], all of which first manner. DINGO is a more recent tool that works similarly to
identify modules co-expressed across the full set of study sam- DiffCoEx by grouping genes based on how differently they be-
ples. These co-expressed modules can then be correlated to pre- have in a particular subset of samples (representing e.g. a par-
defined sample subpopulations representing, for example, ticular condition) from the baseline co-expression determined
disease status or tissue type. from all samples [102]. These are the most likely genes to ex-
WGCNA determines the activity and importance of each plain different phenotypes that are associated with the two dif-
module in each subpopulation of samples (Figure 3A and 3C). ferent networks. Each of the methods detects specific module
For each module, an eigengene is calculated, which is the vector changes by design, but they can also detect modular changes
that best describes the expression behaviour (in a linear fash- that they were not specifically designed for and may outper-
ion) of all genes within this module in the samples included in form other tools in the identification of these changes [130].

Figure 3. Changes in gene co-expression patterns that can occur between samples. Differential co-expression can occur as the presence of a module in only one of the
sample groups (A), as differences in the structure of the module (B) or as differences in the correlation strength between members of the modules (C). Additionally, dif-
ferential co-expression can be detected if one larger interconnected module splits into several smaller ones (D) or if a group of genes changes its correlation partners
[‘gene hopping’ (E)]. If sample groups are not defined before the differential co-expression analysis, or are unknown, biclustering methods can identify modules unique
to a subpopulation of samples by simultaneously classifying the samples into groups in which these modules exist (F).
A number of studies have used differential co-expression detectable by identifying genes that are at the periphery in
network analyses to identify networks unique to specific tissues these modules (Figure 3B). Moreover, some TFs have different
[11] or disease states [134]. The rapid increase in publicly avail- roles in different tissues. These TFs would be expected to be
able RNA-seq data and projects such as GTEx and ENCODE, hub genes that are central to one module under one condition
which generate large-scale RNA-seq profiles, has enabled co- and central to another module in another condition.
expression analysis within and across different tissues [11, 15]. Differentially connected genes are those with different co-
The GTEx project collects and provides expression data from expression partners between two sample groups. These genes
multiple human tissues for the study of gene expression, regu- appear to play a regulatory part in the difference in the pheno-
lation and their relationship to genetic variation [135]. In a study type observed between two groups (Figure 3D) [8–10]. For ex-
comparing RNA-seq data from 35 tissues from the GTEx data ample, one study compared co-expression in mutant cattle with
set, a tissue hierarchy was constructed based on the average increased muscle growth with co-expression in non-mutants,
gene expression in each tissue. Related tissues, such as those using a method similar to DiffCoEx. By identifying the most dif-
from different brain regions, clustered together. This hierarchy ferentially expressed genes and TFs showing the highest differ-
was used to construct a single combined co-expression network ential connection to these genes [10] (Figure 3D), the TF
derived from the tissue-specific co-expression networks—a containing the causal mutation (myostatin) was identified.
meta-network. It was then shown that in tissue-specific net- Interestingly, the Mstn gene, which encodes this TF, hardly
works, TFs with functions specific to that tissue tend to be changed in expression itself, providing an example of how dif-
highly expressed together with tissue-specific genes. These ferential co-expression analysis can uncover biologically im-
genes tend to form a stronger connection with each other than portant findings not revealed by differential expression analysis
with other genes, but remain at the periphery of the network alone.
(thus having low centrality), while the tissue-specific TFs be- Not all methods construct a co-expression network to assess
come more central to that module [11]. Thus, tissue-specific TFs differential expression. GSNCA [103] can be used to identify dif-
could be uncovered by identifying modules with increased co- ferentially co-expressed gene sets, which have to be defined a
expression strength in tissue-specific networks (Figure 3A and priori, between two sample groups. In the first step this method
3C) and by pinpointing the central hubs of these modules. In determines weight vectors for each sample group, from a correl-
contrast, genes that are not TFs but are tissue-specific should be ation network. These weight vectors represent the cross-
correlation of each gene with all the other genes, effectively common in cancer, where different mutations can lead to dif-
summarizing a correlation matrix into a single vector, describ- ferent alterations in co-expression patterns but a similar pheno-
ing a weight for each gene. These weights for the genes repre- type [7]. Biclustering allows researchers to disentangle the
senting a certain gene set are then compared between two mechanisms in the cases where predefining biologically rele-
sample groups, to determine whether the gene set is differen- vant sample groups is difficult. For this purpose, biclustering is
tially co-expressed. more effective than other co-expression analysis methods [7].
Cheng et al. were first to use biclustering in co-expression
Generalized Single Value Decomposition (GSVD) analysis [141], followed by the development and application of
Generalized Single Value Decomposition (GSVD) is a unique many more biclustering approaches (reviewed by Pontes et al.
type of differential co-expression analysis that relies on spectral [106]). The choice of biclustering method depends on the num-

decomposition to identify modules of co-regulated genes. ber of samples and factors such as whether the samples are
Unique to this approach is that it summarizes the expression of species- or tissue-specific and whether the included samples
samples and all genes into a smaller number of variables, aim- constitute disease phenotypes and/or different time points.
ing to explain as much expression variation in as few variables Biclustering methods can be computationally challenging de-
as possible. Here we focus on the summary of gene expression pending on the method used [106]. Methods should be selected
into principal components or ‘genelets’, a term introduced in carefully because different biclustering approaches can have
[104] that can be interpreted as an analogy to co-expressed varying results in the same data set [142].
modules, and which represent the partial expression of mul- Biclustering approaches were recently applied to RNA-seq-
tiple genes. The relative significance of these genelets— based expression data. Analysis of the expression data from
describing the extent to which a signal from the genelet is pre- several developmental stages of worm and fruit fly, by identify-
sent (that is, the extent to which the genelet is expressed) in a ing biclusters containing similar orthologous gene sets unique
data set—can be compared between two data sets. If the signifi- to different developmental stages between the two species, led
cance is similar, the genelet represents a co-expression pattern to the identification of genes with a similar, and thus conserved,
shared between the two data sets, whereas differences in sig- function in development [132]. Biclustering has also been
nificance indicate that the co-expression pattern is unique to applied to single-cell RNA-seq data [5]. Because biclustering
one of the data sets. Higher Order (HO)-GSVD was more recently groups genes and samples simultaneously, it enabled the sim-
developed and uses a similar approach for comparisons be- ultaneous identification of groups of cell types and correspond-
tween more than two data matrices [105]. ing gene modules to reveal 49 different cell types and their
GSVD was first used in 2003 to analyse microarray expres- corresponding cell-type-specific gene modules, results that
sion data from human and budding yeast to identify common were later supported by experimental validation [5]. With the
and unique pheromone and stress response patterns between emergence of single-cell RNA-seq, biclustering methods may be
these two species [104]. HO-GSVD recently proved effective at able to identify cell-type-specific modules that are present in
identifying pathways important for self-renewal of neural pro- diseased but not in healthy cells.
genitors [136]. GSVD was shown to identify patterns unique to Another biclustering method identified miRNAs deregulated
glioblastoma multiforme, a type of brain tumour, which was in breast cancer through their presence in biclusters unique to
useful for prognostic purposes [137]. Similarly, genelets that are cancer samples [7]. These miRNAs have been suggested as
active in normal samples were identified [138]. These genelet markers for diagnosis and treatment response [7]. Biclustering
signals were then removed from the total signal in cancer sam- has also been used to identify tightly co-expressed sets of
ples, revealing a cancer-specific signature [138]. Both of these protein-coding genes unique to subpopulations of cancer pa-
studies demonstrated that signatures unique to the cancer had tients, which could be used to understand patient prognosis
a strong signal for genes duplicated in the cancer [137, 138], as is and to further precision medicine approaches [18, 20]. In an-
common in cancers, suggesting that identified profiles reflect other cancer data set, a three-dimensional clustering method
the oncogenic events in the genome. (triclustering) was used to identify genes co-expressed across
It is not surprising that differential co-expression methods subpopulations of samples and time points [6]. This method ef-
are growing in popularity as the cost of high-quality expression fectively identified several known breast cancer genes in a
data decreases. While these methods have not yet been applied breast cancer cell line by identifying hub genes in triclusters dif-
to RNA-seq data, recent findings from microarray studies make ferentially expressed between cancer samples at early and late
this an exciting prospect. However, because these methods are time points using the eigengene changes between the samples
sensitive to outliers, they require high-quality data. of each tricluster [6].
Differential co-expression without prior grouping Comparison of differential co-expression

An alternative method for detecting differentially expressed
analysis methods
clusters between subpopulations of data is biclustering. If a data While a comprehensive and unbiased comparison of methods
set contains several biologically distinct but unknown sample used in differential co-expression analysis is desirable, the per-
groups, biclustering can identify genes with a similar expression formance of the tools may be situation-dependent, varying be-
pattern in only a sub-set of the samples without the need for tween species, disease states and perhaps even data sets, thus
prior sample classification (Figure 3F). This is particularly useful making it difficult to identify the optimal method in each cir-
when such information is not available, as can be the case for cumstance. An attempt was recently made to compare 10 differ-
large-scale single-cell RNA-seq experiments like those using the ential co-expression algorithms, but concluded that it remains
Drop-seq system [139] or inDrop [140]. difficult to evaluate these owing to the lack of gold standard
In a clinical study it is often possible to predefine groups of gene sets to validate the outcome of these methods [143].
healthy and diseased samples. However, the same disease can Several of the tools described in this review have been com-
manifest through different mechanisms. This is a scenario pared in publications introducing a competing method. DICER
has been argued to perform better than DiffCoEx and CoXpress types of data can help to prioritize genes that may underlie a
[4] based on functional enrichment analysis of differentially ex- phenotype. This can be achieved, for example, using informa-
pressed modules. HO-GSVD outperformed WGCNA and tion describing which genes are TFs, as is done for regulatory
DiffCoEx based on its ability to detect clusters in simulated data predictions by GENIE3 [113]. However, a focus on TFs is rarely
[136]. Although biclustering is a powerful approach, it does not sufficient, and integration of multiple data types is often
necessarily perform better than other network analysis meth- required to increase the accuracy and usefulness of the result-
ods such as WGCNA, as shown by a comparison using different ing networks [13, 147].
tools on simulated data [144]. However, as discussed earlier,
biclustering can be performed without the need for prior sample TF binding site analysis
group classification.
Genome-wide transcription factor binding site (TFBS) analysis

Although many of the tools and methods described in this
review were originally created for microarray data, they are also was introduced in the beginning of this millennium using chro-
applicable to RNA-seq data. There are RNA-seq-specific differ- matin immunoprecipitation followed by microarray analysis,
ential co-expression analysis methods, harnessing the exon- or also known as ChIP-chip [148], which was later replaced by the
isoform-specific expression information or allele-specific ex- more accurate ChIP-seq [149]. These data were used to create a
pression effects, that have been reported to perform better than genome-wide integrated regulatory network from gene expres-
other tools where this information is not considered [52, 145]. sion and TFBS data [150]. Combined analysis of ChIP-chip-based
However, it is unclear whether these differential co-expression TFBSs and expression data initially showed that, in 58% of the
methods also perform better if other methods are supplied with cases, the TFs bound to the promoter region of the gene were in-
the same isoform-specific expression information, which could deed regulated by the corresponding TF [151]. A partial least
be determined before differential co-expression analysis. As a squares approach (a well-known method for analysis of high-
result, it remains difficult to assess whether these new tools dimensional data with several continuous response variables)
perform better than already well-established tools such as was later proposed to identify false positives and distinguish
WGCNA, which can also be used on isoform-specific expression the activation and repression activities of TFs [152]. A more re-
data [50]. cent method harnesses the rapidly increasing availability of
Because the tools described in this review create modules ChIP-seq data in combination with expression data to rank the
based on different criteria, it is also questionable whether the genes bound by a TF, which can be used to prioritize the most
measures used in the comparisons represent desirable proper- likely TF targets [153]. Tools to conduct similar analyses, inte-
ties for all cases. For example, DiffCoEx groups genes based on grating expression and ChIP data, have also been published
their differential co-expression behaviour, whereas WGCNA [154].
identifies modules that are co-expressed in multiple samples
and conditions. In a homogenous data set, DiffCoEx will likely Multilayer integrated networks
detect fewer and smaller modules, indicating that there are not
Independent from the approach used to identify them, network
many differentially co-expressed genes. This will likely lead to
lower enrichment scores when the performance of DiffCoEx is modules can be further investigated for shared eQTL gene tar-
compared with WGCNA on such homogenous data, whereas it gets, TF/miRNA targets or enriched binding motifs [15, 120].
is merely an indication that not many co-expression partner Several computational methods and publicly available data sets
changes occur in the data. By contrast, tools that focus on are available for multi-omics data integration. For example, in-
shared co-expression modules are likely to find strong correl- formation about eQTLs can be acquired from recent large-scale
ation modules with high enrichment scores, which may not be blood-based trans-eQTL meta-analysis [155] or eQTL studies
relevant if the goal is to identify regulatory modules. WGCNA conducted in other tissue types [156]. Transcription factor bind-
has been widely shown to perform well under many different ing sites (TFBSs) can be collected from databases such as
circumstances and for different purposes [54]. However, it re- JASPAR and DeepBind [157], which consist of TF binding motifs
quires information on the sample conditions to assign modules inferred from experimental data. Binding sites can be further
to conditions. If this information is not available (as in large- prioritized by investigating tissue-specific ChIP-seq peaks from
scale single-cell RNA-seq experiments) or if researchers wish to ENCODE [15]. Finally, miRNA–target interactions can be identi-
identify subgroups within the sample groups, biclustering is a fied using several in silico target prediction tools [158, 159] or
more suitable approach. using manually curated databases of experimentally supported
To get a systematic assessment of the performance of differ- target interactions [160–162].
ent tools and methods, projects such as DREAM4 and DREAM5 Combining information from different layers of data may
[146] have been invaluable. These challenged researchers to lead to new biologically interpretable associations in a number
construct regulatory networks from simulated and in vivo of ways. If intra-modular hub genes are TFs or targets of a TF,
benchmark data sets. As these challenges are predefined they this TF is more likely to have a causal role in the phenotype
allow researchers to test their methods/tools in an unbiased under investigation [10]. If multiple Genome-Wide Association
manner. However, these challenges were last posed in 2010 and Study (GWAS) hits exist in the same module, their cumulative
many new methods and tools have been developed since. presence can significantly contribute to disease development
[120, 163, 164]. Differential methylation states of genes within a
co-expression module can elucidate methylation patterns
Integrated network analysis underlying disease [165]. If multiple genes are regulated by the
Experimental validation often focuses on single genes. As these same genetic variant (under a trans-eQTL effect), it may be pos-
experiments are costly and time-consuming, high confidence sible to identify the gene responsible for the alterations of the
predictions of causal genes are of great importance. An analysis network by identifying the cis-eQTL gene driving the trans-eQTL
based solely on co-expression does not (yet) provide this level of effects (Figure 4). This is supported by the fact that genes under
confidence. Therefore, incorporation of information from other trans-regulation of disease-associated genomic variants are

Figure 4. Strategies for integrating multi-omics data with co-expression analyses. Networks are more informative if they are constructed using expression data specific
to the tissue of interest. Genomic variation can be mapped to a co-expression network either by linking suggestive GWAS hits to the genes in the network or by first
identifying genetic variants with an effect on gene expression levels (cis- and trans-eQTLs) and then mapping those to the co-expression network. Additional data
layers may include TFBSs (based on binding motifs or ChIP-seq/ChIP-chip experiments), miRNA target binding sites (based on in silico predictions or experimental tech-
niques) and established protein–protein interactions. A co-expression network can be used to identify modules, hub genes and for predicting the function of unknown
trait-associated genes. Identified modules can be analysed by enrichment analyses to identify overlaying features. Additionally, the research hypothesis can be sup-
ported by additional differential expression, co-expression and methylation analyses that can be performed if respective omics data are available for cases and con-
trols for a corresponding trait. eQTL: expression quantitative trait loci; GWAS: genome-wide association study; OMIM: online Mendelian inheritance in man; miRNA:
microRNA; PPI: protein–protein interaction; TF: transcription factor; TFBS: TF binding site.
sometimes functionally connected with the processes or path- studies [167]. A recently published tool, CoRegNet, allows the in-
ways associated with the corresponding disease. Good ex- tegration of different types of data in a co-expression analysis
amples of this are IFN (interferon)-a and complement pathways by identifying co-operative regulators of genes from different
in which several genes were under trans-regulation of a sys- data types [114]. Another established approach, cMonkey,
temic lupus erythematosus-associated variant, possibly via cis- achieves similar data integration by calculating the joint biclus-
regulation of IKZF1 [155]. The integration of regulatory genetic ter membership probability from different data types by iden-
variant information into co-expression network analysis, with tifying groups of genes that group together in multiple data
cis-eQTLs used as causal anchors, identified TYROBP as the most types [115].
likely causal factor in late-onset Alzheimer disease patients, a
finding supported by the observation that mutations in this
Future prospects
gene are known to cause Nasu-Hakola disease [128]. Lastly,
copy number variation can affect gene expression levels, and In recent years, differential co-expression analyses have been
including such information may help identify and/or explain al- increasingly used to analyse large data sets. This may be attrib-
terations in co-expression network structures present in dis- uted to the decreased costs of large-scale gene expression
eases or traits [138]. profiling, in particular RNA-seq, to increased sample sizes, and
Overall, integration of multiple data types increases the ac- to the greater availability of tissue-specific data from perturb-
curacy of the resulting predictions [13, 147]. For example, mod- ation experiments, which are required for fruitful differential
ules unique to different subtypes of cancer were identified by co-expression analyses [103, 168]. Likewise, biclustering algo-
integrating tumour genome sequences with gene networks rithms have benefitted from larger sample sizes and higher
[166], and these modules may be useful for prognosis and iden- data quality, as shown by the identification of co-expressed
tification of putative targets for personalized medicine-based modules unique to cancer subtypes [18, 20]. The usefulness of
treatments. A number of tools, described earlier in this review, biclustering on single-cell RNA-seq data has been demonstrated
can be used for differential co-expression analysis, but can also by the classification of different cell types and by the identifica-
be applied to other data types. In the initial DINGO publication, tion of clusters of genes uniquely co-expressed in specific cell
the authors conducted a combined analysis on mRNA expres- types [5]. We expect these approaches to be more widely applied
sion, DNA copy number variation and methylation data. By in the future, as they benefit from an increase in RNA-seq data
overlaying the differential networks of each data type and iden- quantity and quality, which will allow for more accurate identi-
tifying edges present in all of them, a number of genes from the fication of tissue-specific and cell-type-specific disease-related
PI3K pathway were identified as important players in glioblast- modules and regulators.
oma multiforme patients [102]. This pathway is an already-es- Large-scale single-cell sequencing technology is increasingly
tablished therapeutic target, supporting the notion that this is used and the first co-expression studies using such techniques
an effective approach for identifying relevant targets for disease have uncovered cell-type-specific co-expression modules that
would have gone undetected in multi-cell-type co-expression data, and the tools that exist mostly integrate only two layers of
analyses [5, 12]. Because the latter represent the aggregated sig- omics data [177]. Integrated network analyses come with add-
nals of multiple cell types, they usually cannot detect alter- itional mathematical challenges, and best practices are far from
ations in cell subpopulations between different experimental established. Further research on this topic is of great interest to
groups. This is supported by the observation that the expression the research community, as it will allow a better understanding
of cell cycle genes associated with ageing decreased in the ana- of regulatory mechanisms that can explain co-expression pat-
lysis of non-cell-type-specific data [169]. However, data from terns and disease mechanisms. A better understanding of these
single-cell experiments revealed that this observation was disease mechanisms and corresponding co-expression patterns
caused by a decreased proportion of the G1/S cells that highly will facilitate the identification of appropriate targets for inter-
express cell cycle genes rather than by altered expression across vention studies.

the whole cell population [170].
An additional prospect is the detection of mutations from
RNA-seq data [171]. As mutations accumulate with age in differ- Key Points
ent cells, these can be used to identify the origin of the cell.
• RNA-seq-based co-expression analysis can be used to
Mutation accumulation has been used to study cancer develop-
ment and the origin of metastases [172]. In large-scale single- assign putative functions to non-coding RNAs and to
cell RNA-seq experiments, mutations could be used to separate identify candidates for roles in disease.
• In co-expression networks, hub gene identification has
cells based on their origin, or to group cells based on the muta-
tions they harbour [173]. Cells harbouring the same mutations a limited power for identifying targets for follow-up
can be investigated for co-expression patterns, and modules studies; yet, this can be enhanced by integrated net-
unique to cells with a specific mutation may be detected. This work analyses, which may incorporate GWAS hits,
may allow the direct linking of mutations to expression mod- eQTLs, TFBSs and other data layers.
ules, with the limitation that only mutations in coding regions • Differential co-expression analyses can reveal genes
are detectable in RNA-seq data. that have different co-expression partners between
Although there are many exciting new possibilities with healthy and disease state and can help to uncover
single-cell RNA-seq data, important challenges remain. Typically, regulators underlying disease and other phenotypes.
a low number of reads per cell are sequenced and then the signal • Methods such as biclustering and Generalised Single
from multiple cells of the same type is aggregated to acquire a Value Decomposition (GSVD) allow the identification
cell-type-specific gene-expression profile. It is hard to acquire of signals/modules unique to specific cancer subtypes,
sufficient data for rarer cell populations, such as stem cells, and which may serve a purpose in prognosis and for preci-
this is currently limiting analyses on these cell types. sion medicine.
Additionally, the low number of reads per cell leads to sparse ex-
pression matrixes to which normalization methods currently
Acknowledgements
used in canonical RNA-seq analyses are not attuned. These nor-
malization methods often also assume that the majority of genes We thank the members of the Integrative Genomics of
do not change in expression between different samples, which is Ageing Group for useful suggestions and discussions. We
not necessarily the case in single-cell RNA-seq owing to variation are grateful for the assistance of Aoife Doherty, Monique
in expression across different cells. This is further exacerbated by van der Wijst and Kate McIntyre with the writing and edit-
the difficulty in obtaining high-quality RNA from single cells. ing of this manuscript. Additionally, we thank the members
These and other issues are further discussed in [174]. of FrankeSwertzLab at University of Groningen/University
In addition to the normalization issues that occur in single- Medical Center Groningen for their support and suggestions.
cell RNA-seq, the optimal method for normalizing bulk RNA-seq Finally, we thank the Institute of Integrative Biology at the
data is also still not clear. The widely used Fragment/Reads Per University of Liverpool for supporting the PhD studies of
Kilobase Million (FPKM) normalization has been debated [58]
Sipko van Dam.
and although alternatives have and are being created, each
method has its limitations. Additionally, from our experience,
the use of different mapping tools can in some cases lead to dif- Funding
ferent results. Although some comparisons between different
UK Biotechnology and Biological Sciences Research Council
tools and methods have been made [175], a large-scale compari-
grant (BB/K016741/1 to J.P.M.); European Research Council
son, using e.g. public data, would identify such cases and define
ERC Starting Grant (637640 to L.F.); and the Organization for
best practices for pursuing each research question.
Scientific Research (917.14.374 to L.F.).
With the increased availability of different data types such
as RNA-seq, genome sequences, ChIP-seq, methylome and
proteome data, it will become possible to integrate these data Glossary
sets to more accurately predict regulatory genes. Projects from
large consortia like GTEx [156], the Epigenome Roadmap [176]
Betweenness centrality
and ENCODE [15] are already generating data from multiple- This measure is used to describe the centrality (and there-
omics levels that facilitate these integrated analyses. To iden- fore relevance) of a node in the network by counting the
tify regulatory relationships, perturbation data are preferable, number of shortest paths between any other pair of nodes
as canonical data cannot distinguish between true and false going through this node (Figure 2).
positives in regulatory relationships [129, 168]. Furthermore, Biclustering
regulatory relationships can be highly cell-type-, tissue- or A clustering method that allows clustering based on two di-
developmental-stage-specific [129]. Only a handful of tools and mensions simultaneously, e.g. genes and different
methods are currently available to investigate multi-omics treatments.
ChIP-chip Module
This method identifies TFBSs by immunoprecipitation of A group of co-expressed genes that form a sub-network in
the TF together with bound DNA fragments (chromatin the larger network, usually defined by applying clustering
immunoprecipitation—ChIP). A DNA microarray is subse- algorithms on a co-expression network or directly on ex-
quently used to identify the sequences where the corres- pression profiles.
ponding TF is bound. Mutual information
ChIP-seq The measure of dependence between two otherwise unre-
This method uses the same approach as ChIP-chip, but lated variables.
using RNA-seq rather than microarray to identify TFBSs. Network robustness
Clustering A measure of how resistant a network is to the removal of sin-

A statistical method of grouping variables. In the case of gle nodes, assessed by the effect of removal on the connectiv-
gene co-expression, genes that have similar expression pat- ity of the network. The connectivity indicates how many
terns across multiple samples. nodes need to be removed to disconnect part of a network.
Co-expression partner Regulatory gene
A gene whose expression shows a similar pattern across dif- A gene that regulates the expression of other genes.
ferent samples to that of a gene of interest. Subpopulation-specific modules
Co-expression network A co-expression module that only exists in a subset of the
A network that describes which genes have a tendency to samples analysed.
show a coordinated expression pattern across a group of Transcript
samples. In these networks, each node represents a gene A single-stranded RNA molecule resulting from the tran-
and each edge represents the presence and/or the strength scription of a gene.
of the co-expression relationship. Triclustering
Eigengene A clustering method that allows clustering in three dimen-
A vector that best describes the expression changes of a sions simultaneously, e.g. genes, different treatments and
module between different samples. This describes the par- time points.
tial expression of each gene (albeit to a different extent for
each gene) in a linear fashion. References
Gene
1. Zhao Y, Li H, Fang S, et al. NONCODE 2016: an informative
A DNA sequence that can be transcribed into a transcript. In
and valuable data source of long non-coding RNAs. Nucleic
the case of protein coding genes, this transcript can be
Acids Res 2016;44:D203–8.
translated into a protein. Proteins are the building blocks of
2. van Dam S, Craig T, de Magalhaes JP. GeneFriends: a human
our body. Non-coding genes are transcribed but do not en-
RNA-seq-based gene and transcript co-expression database.
code proteins.
Nucleic Acids Res 2015;43:D1124–32.
Genelet 3. Carpenter AE, Sabatini DM. Systematic genome-wide
Similar to the eigengene; a vector that represents the partial screens of gene function. Nat Rev Genet 2004;5:11–22.
expression of multiple genes, but calculated using a different 4. Amar D, Safer H, Shamir R. Dissection of regulatory net-
method. Unlike an eigengene, a genelet does not necessarily works that are altered in disease via differential co-expres-
best explain the expression variability present in the data. sion. PLoS Comput Biol 2013;9:e1002955.
Generalized single value decomposition 5. Zeisel A, Munoz-Manchado AB, Codeluppi S, et al. Brain
A type of differential co-expression analysis that compares the structure. Cell types in the mouse cortex and hippocampus
strength of gene expression signatures existing in one group of revealed by single-cell RNA-seq. Science 2015;347:1138–42.
samples with the strength for this signal in another group. 6. Bhar A, Haubrock M, Mukhopadhyay A, et al. Coexpression
Hierarchal clustering and coregulation analysis of time-series gene expression
A clustering method that creates a hierarchical tree based data in estrogen-induced breast cancer cell. Algorithms Mol
on the distance between the nodes or genes in the network. Biol 2013;8:9.
Nodes that are close to each other in the network are part of 7. Fiannaca A, La Rosa M, La Paglia L, et al. Analysis of miRNA
the same branch. Clusters can be defined by cutting the expression profiles in breast cancer using biclustering. Bmc
branches at a certain height. Bioinformatics 2015;16 Suppl 4:S7.
8. Hu R, Qiu X, Glazko G, et al. Detecting intergene correlation
Intra-modular hub
changes in microarray analysis: a new approach to gene se-
A gene with high centrality within a module. These tend to
lection. Bmc Bioinformatics 2009;10:20.
have high biological relevance.
9. Kostka D, Spang R. Finding disease specific alterations in the
Inter-modular hub
co-expression of genes. Bioinformatics 2004;20 (Suppl 1):i194–9.
A gene with high centrality in a network.
10. Hudson NJ, Reverter A, Dalrymple BP. A differential wiring
k-means clustering analysis of expression data correctly identifies the gene con-
A clustering method where a predefined number of clusters taining the causal mutation. PLoS Comput Biol 2009;5:e1000382.
are constructed so that each gene is assigned to the cluster 11. Pierson E, Consortium GT, Koller D, et al. Sharing and specifi-
with the closest mean. city of co-expression networks across 35 human tissues.
Microarray PLoS Comput Biol 2015;11:e1004220.
A platform for quantifying gene expression that assays 12. Xue Z, Huang K, Cai C, et al. Genetic programs in human and
mRNA molecules based on their hybridization to probes pre- mouse early embryos revealed by single-cell RNA sequenc-
sent on an array, typically a glass slide. ing. Nature 2013;500:593–7.
13. Glass K, Huttenhower C, Quackenbush J, et al. Passing mes- neuronal activity-dependent genes in autism. Nat Commun
sages between biological networks to refine predicted inter- 2014;5:5748.
actions. PLoS One 2013;8:e64832. 35. de Magalhaes JP, Finch CE, Janssens G. Next-generation
14. De Smet R, Marchal K. Advantages and limitations of current sequencing in aging research: emerging applications, prob-
network inference methods. Nat Rev Microbiol 2010;8:717–29. lems, pitfalls and possible solutions. Ageing Res Rev
15. Yue F, Cheng Y, Breschi A, et al. A comparative encyclopedia 2010;9:315–23.
of DNA elements in the mouse genome. Nature 36. Chen J, Bardes EE, Aronow BJ, et al. ToppGene Suite for gene
2014;515:355–64. list enrichment analysis and candidate gene prioritization.
16. Goh KI, Cusick ME, Valle D, et al. The human disease net- Nucleic Acids Res 2009;37:W305–11.
work. Proc Natl Acad Sci USA 2007;104:8685–90. 37. Mason MJ, Fan G, Plath K, et al. Signed weighted gene co-

17. Lage K, Hansen NT, Karlberg EO, et al. A large-scale analysis expression network analysis of transcriptional regulation in
of tissue-specific pathology and gene expression of human murine embryonic stem cells. BMC Genomics 2009;10:327.
disease genes and complexes. Proc Natl Acad Sci USA 38. Lim LP, Lau NC, Garrett-Engele P, et al. Microarray analysis
2008;105:20870–5. shows that some microRNAs downregulate large numbers
18. Wang YK, Print CG, Crampin EJ. Biclustering reveals breast of target mRNAs. Nature 2005;433:769–73.
cancer tumour subgroups with common clinical features 39. Hrdlickova B, de Almeida RC, Borek Z, et al. Genetic variation
and improves prediction of disease recurrence. BMC in the non-coding genome: Involvement of micro-RNAs and
Genomics 2013;14:102. long non-coding RNAs in disease. Biochim Biophys Acta
19. Yang Y, Han L, Yuan Y, et al. Gene co-expression network 2014;1842:1910–22.
analysis reveals common system-level properties of prog- 40. Zhang B, Horvath S. A general framework for weighted gene
nostic genes across cancer types. Nat Commun 2014;5:3231. co-expression network analysis. Stat Appl Genet Mol Biol
20. Ahmadi Adl A, Qian X. Tumor stratification by a novel 2005;4:Article17.
graph-regularized bi-clique finding algorithm. Comput Biol 41. Khalil AM, Guttman M, Huarte M, et al. Many human large
Chem 2015;57:3–11. intergenic noncoding RNAs associate with chromatin-
21. Ballouz S, Verleyen W, Gillis J. Guidance for RNA-seq co-ex- modifying complexes and affect gene expression. Proc Natl
pression network construction and analysis: safety in num- Acad Sci USA 2009;106:11667–72.
bers. Bioinformatics 2015;31:2123–30. 42. Garcia-Blanco MA, Baraniak AP, Lasda EL. Alternative splic-
22. Steuer R, Kurths J, Daub CO, et al. The mutual information: ing in disease and therapy. Nat Biotechnol 2004;22:535–46.
Detecting and evaluating dependencies between variables. 43. Jin G, Sun J, Isaacs SD, et al. Human polymorphisms at long
Bioinformatics 2002;18:S231–40. non-coding RNAs (lncRNAs) and association with prostate
23. Margolin AA, Nemenman I, Basso K, et al. ARACNE: an algo- cancer risk. Carcinogenesis 2011;32:1655–9.
rithm for the reconstruction of gene regulatory networks in 44. Ge W, Ma X, Li X, et al. B7-H1 up-regulation on dendritic-like
a mammalian cellular context. Bmc Bioinformatics 2006;7 leukemia cells suppresses T cell immune function through
(Suppl 1):S7. modulation of IL-10/IL-12 production and generation of Treg
24. Butte AJ, Kohane IS. Mutual information relevance net- cells. Leuk Res 2009;33:948–57.
works: functional genomic clustering using pairwise en- 45. Sekhon RS, Briskine R, Hirsch CN, et al. Maize gene atlas de-
tropy measurements. Pac Symp Biocomput 2000;418–29. veloped by RNA sequencing and comparative evaluation of
25. Guttman M, Donaghey J, Carey BW, et al. lincRNAs act in the transcriptomes based on RNA sequencing and microarrays.
circuitry controlling pluripotency and differentiation. Nature PLoS One 2013;8:e61005.
2011;477:295–U260. 46. Richard H, Schulz MH, Sultan M, et al. Prediction of alterna-
26. Ala U, Piro RM, Grassi E, et al. Prediction of human disease tive isoforms from exon expression levels in RNA-Seq ex-
genes by human-mouse conserved coexpression analysis. periments. Nucleic Acids Res 2010;38:
PLoS Comput Biol 2008;4:e1000043. 47. Wood SH, Craig T, Li Y, et al. Whole transcriptome sequenc-
27. van Someren EP, Vaes BL, Steegenga WT, et al. Least absolute ing of the aging rat brain reveals dynamic RNA changes in
regression network analysis of the murine osteoblast differ- the dark matter of the genome. Age (Dordr) 2013;35:763–76.
entiation network. Bioinformatics 2006;22:477–84. 48. Yang X, Coulombe-Huntington J, Kang S, et al. Widespread
28. Friedman N, Linial M, Nachman I, et al. Using Bayesian net- Expansion of Protein Interaction Capabilities by Alternative
works to analyze expression data. J Comput Biol 2000;7: Splicing. Cell 2016;164:805–17.
601–20. 49. Kelemen O, Convertini P, Zhang Z, et al. Function of alterna-
29. D’Haeseleer P, Liang S, Somogyi R. Genetic network infer- tive splicing. Gene 2013;514:1–30.
ence: from co-expression clustering to reverse engineering. 50. Iancu OD, Colville A, Oberbeck D, et al. Cosplicing network
Bioinformatics 2000;16:707–26. analysis of mammalian brain RNA-Seq data utilizing
30. Kumari S, Nie J, Chen HS, et al. Evaluation of gene associ- WGCNA and Mantel correlations. Front Genet 2015;6:174.
ation methods for coexpression network construction and 51. Dai C, Li W, Liu J, et al. Integrating many co-splicing net-
biological knowledge discovery. PLoS One 2012;7:e0050411. works to reconstruct splicing regulatory modules. BMC Syst
31. Senbabaoglu Y, Sumer SO, Sanchez-Vega F, et al. A multi- Biol 2012;6 (Suppl 1):S17.
method approach for proteomic network inference in 11 52. Yalamanchili HK, Li Z, Wang P, et al. SpliceNet: recovering
human cancers. PLoS Comput Biol 2016;12:e1004765. splicing isoform-specific differential gene networks from
32. Albert R, Barabasi AL. Statistical mechanics of complex net- RNA-Seq data of normal and diseased samples. Nucleic Acids
works. Rev Modern Phys 2002;74:47–97. Res 2014;42:e121.
33. D’Haeseleer P. How does gene expression clustering work? 53. Cunningham F, Amode MR, Barrell D, et al. Ensembl 2015.
Nat Biotechnol 2005;23:1499–501. Nucleic Acids Res 2015;43:D662–9.
34. Gupta S, Ellis SE, Ashar FN, et al. Transcriptome analysis re- 54. Langfelder P, Horvath S. WGCNA: an R package for weighted
veals dysregulation of innate immune response genes and correlation network analysis. BMC Bioinformatics 2008;9:559.
55. Conesa A, Madrigal P, Tarazona S, et al. A survey of best prac- 76. Zhao W, Langfelder P, Fuller T, et al. Weighted gene coex-
tices for RNA-seq data analysis. Genome Biol 2016;17:13. pression network analysis: state of the art. J Biopharm Stat
56. Li J, Bushel PR. EPIG-Seq: extracting patterns and identifying 2010;20:281–300.
co-expressed genes from RNA-Seq data. BMC Genomics 77. Rodius S, Androsova G, Gotz L, et al. Analysis of the dynamic
2016;17:255. co-expression network of heart regeneration in the zebra-
57. Bacher R, Kendziorski C. Design and computational analysis fish. Sci Rep 2016;6:26822.
of single-cell RNA-sequencing experiments. Genome Biol 78. Gaiteri C, Ding Y, French B, et al. Beyond modules and hubs:
2016;17:63. the potential of gene coexpression networks for investigat-
58. Dillies MA, Rau A, Aubert J, et al. A comprehensive evalu- ing molecular mechanisms of complex brain disorders.
ation of normalization methods for Illumina high- Genes Brain Behav 2014;13:13–24.

throughput RNA sequencing data analysis. Brief Bioinform 79. Dewey FE, Perez MV, Wheeler MT, et al. Gene coexpression
2013;14:671–83. network topology of cardiac development, hypertrophy, and
59. Li S, Tighe SW, Nicolet CM, et al. Multi-platform assessment failure. Circ Cardiovasc Genet 2011;4:26–35.
of transcriptome profiling using RNA-seq in the ABRF next- 80. Rosenkrantz JT, Aarts H, Abee T, et al. Non-essential genes
generation sequencing study. Nat Biotechnol 2014;32:915–25. form the hubs of genome scale protein function and envir-
60. Okamura Y, Aoki Y, Obayashi T, et al. COXPRESdb in 2015: onmental gene expression networks in Salmonella enterica
coexpression database for animal species by DNA- serovar Typhimurium. BMC Microbiol 2013;13:294.
microarray and RNAseq-based expression data with mul- 81. Andrews S. FastQC: A Quality Control Tool for High Throughput
tiple quality assessment systems. Nucleic Acids Res 2015;43: Sequence Data. 2010. http://www.bioinformatics.babraham.
D82–6. ac.uk/projects/fastqc.
61. Giorgi FM, Del Fabbro C, Licausi F. Comparative study of 82. Wang L, Wang S, Li W. RSeQC: quality control of RNA-seq ex-
RNA-seq- and microarray-derived coexpression networks in periments. Bioinformatics 2012;28:2184–5.
Arabidopsis thaliana. Bioinformatics 2013;29:717–24. 83. Hartley SW, Mullikin JC. QoRTs: a comprehensive toolset for
62. Efron BTR, Monographs on Statistics and Applied Probability: An quality control and data processing of RNA-Seq experi-
Introduction to the Bootstrap. New York: Chapman and Hall, ments. BMC Bioinformatics 2015;16:224.
1993, 57. 84. Trapnell C, Pachter L, Salzberg SL. TopHat: discovering splice
63. Langfelder P, Horvath S. Eigengene networks for studying junctions with RNA-Seq. Bioinformatics 2009;25:1105–11.
the relationships between co-expression modules. BMC Syst 85. Dobin A, Davis CA, Schlesinger F, et al. STAR: ultrafast uni-
Biol 2007;1:54. versal RNA-seq aligner. Bioinformatics 2013;29:15–21.
64. Chen YQ, Zhu J, Lum PY, et al. Variations in DNA elucidate 86. Kim D, Langmead B, Salzberg SL. HISAT: a fast spliced
molecular networks that cause disease. Nature 2008;452: aligner with low memory requirements. Nat Methods
429–35. 2015;12:357–60.
65. Langfelder P, Mischel PS, Horvath S. When is hub gene selec- 87. Li H, Durbin R. Fast and accurate short read alignment with
tion better than standard meta-analysis? PLoS One Burrows-Wheeler transform. Bioinformatics 2009;25:1754–60.
2013;8:e0061505. 88. Bray N, Pimentel H, Melsted P, et al. Near-optimal RNA-Seq
66. Kogelman LJ, Cirera S, Zhernakova DV, et al. Identification of Quantification, 2015.
co-expression gene networks, regulatory genes and path- 89. Patro R, Duggal G, Kingsford C, Salmon: Accurate, Versatile and
ways for obesity based on adipose tissue RNA sequencing in Ultrafast Quantification from RNA-seq Data using Lightweight-
a porcine model. BMC Med Genomics 2014;7:57. Alignment, 2015.
67. Lee JH, Gao C, Peng GD, et al. Analysis of transcriptome com- 90. Anders S, Pyl PT, Huber W. HTSeq–a Python framework to
plexity through RNA sequencing in normal and failing mur- work with high-throughput sequencing data. Bioinformatics
ine hearts. Circ Res 2011;109:1332–41. 2015;31:166–9.
68. Langfelder P, Luo R, Oldham MC, et al. Is my network module 91. Liao Y, Smyth GK, Shi W. featureCounts: an efficient general
preserved and reproducible? PLoS Comput Biol 2011;7: purpose program for assigning sequence reads to genomic
e1001057. features. Bioinformatics 2014;30:923–30.
69. Albert R, Jeong H, Barabasi AL. Error and attack tolerance of 92. Trapnell C, Williams BA, Pertea G, et al. Transcript assembly
complex networks. Nature 2000;406:378–82. and quantification by RNA-Seq reveals unannotated tran-
70. Freeman LC. Centrality in social networks conceptual clarifi- scripts and isoform switching during cell differentiation.
cation. Soc Netw 1979;1:215–39. Nat Biotechnol 2010;28:511–5.
71. Chou WC, Cheng AL, Brotto M, et al. Visual gene-network 93. Li B, Dewey CN. RSEM: accurate transcript quantification
analysis reveals the cancer gene co-expression in human from RNA-Seq data with or without a reference genome.
endometrial cancer. BMC Genomics 2014;15:300. BMC Bioinformatics 2011;12:323.
72. Oh EY, Christensen SM, Ghanta S, et al. Extensive rewiring of 94. Robinson MD, Oshlack A. A scaling normalization method
epithelial-stromal co-expression networks in breast cancer. for differential expression analysis of RNA-seq data. Genome
Genome Biol 2015;16:128. Biol 2010;11:R25.
73. Keller MP, Choi Y, Wang P, et al. A gene expression network 95. Sohn MB, Du R, An L. A robust approach for identifying dif-
model of type 2 diabetes links cell cycle regulation in islets ferentially abundant features in metagenomic samples.
with diabetes susceptibility. Genome Res 2008;18:706–16. Bioinformatics 2015;31:2269–75.
74. Presson AP, Sobel EM, Papp JC, et al. Integrated weighted 96. Love MI, Huber W, Anders S. Moderated estimation of fold
gene co-expression network analysis with an application to change and dispersion for RNA-seq data with DESeq2.
chronic fatigue syndrome. BMC Syst Biol 2009;2: Genome Biol 2014;15:550.
75. Voineagu I, Wang XC, Johnston P, et al. Transcriptomic ana- 97. Ritchie ME, Phipson B, Wu D, et al. limma powers differential
lysis of autistic brain reveals convergent molecular path- expression analyses for RNA-sequencing and microarray
ology. Nature 2011;474:380. þ. studies. Nucleic Acids Res 2015;43:e47.
98. Leek JT. svaseq: removing batch effects and other unwanted 119. Zimmermann P, Hirsch-Hoffmann M, Hennig L, et al.
noise from sequencing data. Nucleic Acids Res 2014;42. GENEVESTIGATOR. Arabidopsis microarray database and
99. Johnson WE, Li C, Rabinovic A. Adjusting batch effects in analysis toolbox. Plant Physiol 2004;136:2621–32.
microarray expression data using empirical Bayes methods. 120. Greene CS, Krishnan A, Wong AK, et al. Understanding
Biostatistics 2007;8:118–27. multicellular function and disease with human tissue-
100. Tesson BM, Breitling R, Jansen RC. DiffCoEx: a simple and specific networks. Nat Genet 2015;47:569–76.
sensitive method to find differentially coexpressed gene 121. Singer GAC, Lloyd AT, Huminiecki LB, et al. Clusters of co-
modules. BMC Bioinformatics 2010;11:497. expressed genes in mammalian genomes are conserved by
101. Watson M. CoXpress: differential co-expression in gene ex- natural selection. Mol Biol Evol 2005;22:767–75.
pression data. BMC Bioinformatics 2006;7:509. 122. Segal E, Friedman N, Koller D, et al. A module map showing

102. Ha MJ, Baladandayuthapani V, Do KA. DINGO: differential conditional activity of expression modules in cancer. Nat
network analysis in genomics. Bioinformatics 2015;31: Genet 2004;36:1090–8.
3413–20. 123. Torkamani A, Dean B, Schork NJ, et al. Coexpression network
103. Rahmatallah Y, Emmert-Streib F, Glazko G. Gene Sets Net analysis of neural tissue reveals perturbations in develop-
Correlations Analysis (GSNCA): a multivariate differential mental processes in schizophrenia. Genome Res 2010;20:
coexpression test for gene sets. Bioinformatics 2014;30:360–8. 403–12.
104. Alter O, Brown PO, Botstein D. Generalized singular value 124. Franke L, van Bakel H, Fokkens L, et al. Reconstruction of a
decomposition for comparative analysis of genome-scale functional human gene network, with an application for pri-
expression data sets of two different organisms. Proc Natl oritizing positional candidate genes. Am J Hum Genet
Acad Sci USA 2003;100:3351–6. 2006;78:1011–25.
105. Ponnapalli SP, Saunders MA, Van Loan CF, et al. A higher- 125. McCarroll SA, Murphy CT, Zou S, et al. Comparing genomic
order generalized singular value decomposition for expression patterns across species identifies shared tran-
comparison of global mRNA expression from multiple or- scriptional profile in aging. Nat Genet 2004;36:197–204.
ganisms. PLoS One 2011;6:e28072. 126. van Dam S, Cordeiro R, Craig T, et al. GeneFriends: an online
106. Pontes B, Giraldez R, Aguilar-Ruiz JS. Biclustering on expres- co-expression analysis tool to identify novel gene targets for
sion data: a review. J Biomed Inform 2015;57:163–80. aging and complex diseases. BMC Genomics 2012;13:535.
107. Huang DW, Sherman BT, Lempicki RA. Systematic and inte- 127. Gillis J, Pavlidis P. “Guilt by association” is the exception ra-
grative analysis of large gene lists using DAVID bioinfor- ther than the rule in gene networks. PLoS Comput Biol
matics resources. Nat Protoc 2009;4:44–57. 2012;8:e1002444.
108. Mi HY, Muruganujan A, Casagrande JT, et al. Large-scale 128. Zhang B, Gaiteri C, Bodea LG, et al. Integrated systems ap-
gene function analysis with the PANTHER classification sys- proach identifies genetic nodes and networks in late-onset
tem. Nat Protoc 2013;8:1551–66. Alzheimer’s disease. Cell 2013;153:707–20.
109. Reimand J, Arak T, Vilo J. g:Profiler–a web server for func- 129. Djordjevic D, Yang A, Zadoorian A, et al. How difficult is in-
tional interpretation of gene lists (2011 update). Nucleic Acids ference of mammalian causal gene regulatory networks?
Res 2011;39:W307–15. PLoS One 2014;9:e111661.
110. Yu G, Wang LG, Han Y, et al. clusterProfiler: an R package for 130. Allen JD, Xie Y, Chen M, et al. Comparing statistical methods
comparing biological themes among gene clusters. Omics for constructing large scale gene networks. PLoS One
2012;16:284–7. 2012;7:e29348.
111. Chen EY, Tan CM, Kou Y, et al. Enrichr: interactive and col- 131. Salgado H, Gama-Castro S, Peralta-Gil M, et al. RegulonDB
laborative HTML5 gene list enrichment analysis tool. BMC (version 5.0): Escherichia coli K-12 transcriptional regulatory
Bioinformatics 2013;14:128. network, operon organization, and growth conditions.
112. Chavez Montes RA, Coello G, Gonzalez-Aguilera KL, et al. Nucleic Acids Res 2006;34:D394–7.
ARACNe-based inference, using curated microarray data, of 132. Gao Q, Ho C, Jia Y, et al. Biclustering of linear patterns in
Arabidopsis thaliana root transcriptional regulatory net- gene expression data. J Comput Biol 2012;19:619–31.
works. BMC Plant Biol 2014;14:97. 133. Monaco G, van Dam S, Casal Novo Ribeiro JL, et al. A com-
113. Huynh-Thu VA, Irrthum A, Wehenkel L, et al. Inferring regu- parison of human and mouse gene co-expression networks
latory networks from expression data using tree-based reveals conservation and divergence at the tissue, pathway
methods. PLoS One 2010;5:e0012776. and disease levels. BMC Evol Biol 2015;15:259.
114. Nicolle R, Radvanyi F, Elati M. CoRegNet: reconstruction and 134. Anglani R, Creanza TM, Liuzzi VC, et al. Loss of connectivity
integrated analysis of co-regulatory networks. Bioinformatics in cancer co-expression networks. PLoS One 2014;9:e87075.
2015;31:3066–8. 135. Mele M, Ferreira PG, Reverter F, et al. Human genomics. The
115. Reiss DJ, Plaisier CL, Wu WJ, et al. cMonkey2: Automated, human transcriptome across tissues and individuals. Science
systematic, integrated detection of co-regulated gene mod- 2015;348:660–5.
ules for any organism. Nucleic Acids Res 2015;43:e87. 136. Xiao X, Moreno-Moral A, Rotival M, et al. Multi-tissue ana-
116. Shannon P, Markiel A, Ozier O, et al. Cytoscape: a software lysis of co-expression networks by higher-order generalized
environment for integrated models of biomolecular inter- singular value decomposition identifies functionally coher-
action networks. Genome Res 2003;13:2498–504. ent transcriptional modules. PLoS Genet 2014;10:e1004006.
117. Theocharidis A, van Dongen S, Enright AJ, et al. Network 137. Lee CH, Alpert BO, Sankaranarayanan P, et al. GSVD com-
visualization and analysis of gene expression data using parison of patient-matched normal and tumor aCGH pro-
BioLayout Express(3D). Nat Protoc 2009;4:1535–50. files reveals global copy-number alterations predicting
118. Vlasblom J, Zuberi K, Rodriguez H, et al. Novel function dis- glioblastoma multiforme survival. PLoS One 2012;7:e30098.
covery with GeneMANIA: a new integrated resource for gene 138. Fehrmann RS, Karjalainen JM, Krajewska M, et al. Gene ex-
function prediction in Escherichia coli. Bioinformatics 2015;31: pression analysis identifies global gene dosage sensitivity in
306–10. cancer. Nat Genet 2015;47:115–25.
139. Macosko EZ, Basu A, Satija R, et al. Highly parallel genome- 159. John B, Enright AJ, Aravin A, et al. Human MicroRNA targets.
wide expression profiling of individual cells using nanoliter PLoS Biol 2004;2:e363.
droplets. Cell 2015;161:1202–14. 160. Vlachos IS, Paraskevopoulou MD, Karagkouni D, et al.
140. Klein AM, Mazutis L, Akartuna I, et al. Droplet barcoding for DIANA-TarBase v7.0: indexing more than half a million ex-
single-cell transcriptomics applied to embryonic stem cells. perimentally supported miRNA:mRNA interactions. Nucleic
Cell 2015;161:1187–201. Acids Res 2015;43:D153–9.
141. Cheng Y, Church GM. Biclustering of expression data. Proc 161. Chou CH, Chang NW, Shrestha S, et al. miRTarBase 2016: up-
Int Conf Intell Syst Mol Biol 2000;8:93–103. dates to the experimentally validated miRNA-target inter-
142. Oghabian A, Kilpinen S, Hautaniemi S, et al. Biclustering actions database. Nucleic Acids Res 2016;44:D239–47.
methods: biological relevance and application in gene ex- 162. Kozomara A, Griffiths-Jones S. miRBase: annotating high

pression analysis. PLoS One 2014;9:e90801. confidence microRNAs using deep sequencing data. Nucleic
143. Lichtblau Y, Zimmermann K, Haldemann B, et al. Acids Res 2014;42:D68–73.
Comparative assessment of differential network analysis 163. Naukkarinen J, Surakka I, Pietilainen KH, et al. Use of
methods. Brief Bioinform 2016, doi: 10.1093/bib/bbw061. genome-wide expression data to mine the “Gray Zone” of
144. Bo V, Curtis T, Lysenko A, et al. Discovering study-specific GWA studies leads to novel candidate obesity genes. PLoS
gene regulatory networks. PLoS One 2014;9:e106524. Genet 2010;6:e1000976.
145. Hong S, Chen X, Jin L, et al. Canonical correlation analysis for 164. Corradin O, Saiakhova A, Akhtar-Zaidi B, et al.
RNA-seq co-expression networks. Nucleic Acids Res 2013;41:e95. Combinatorial effects of multiple enhancer variants in link-
146. Marbach D, Costello JC, Kuffner R, et al. Wisdom of crowds age disequilibrium dictate levels of gene expression to con-
for robust gene network inference. Nat Methods 2012;9:
fer susceptibility to common traits. Genome Res 2014;24:1–13.
796–804.
165. Fernandez-Jimenez N, Castellanos-Rubio A, Plaza-Izurieta L,
147. Hecker M, Lambeck S, Toepfer S, et al. Gene regulatory net-
et al. Coregulation and modulation of NFkappaB-related
work inference: data integration in dynamic models-a re-
genes in celiac disease: uncovered aspects of gut mucosal
view. Biosystems 2009;96:86–103.
inflammation. Hum Mol Genet 2014;23:1298–310.
148. Buck MJ, Lieb JD. ChIP-chip: considerations for the design,
166. Hofree M, Shen JP, Carter H, et al. Network-based stratifica-
analysis, and application of genome-wide chromatin immu-
tion of tumor mutations. Nat Methods 2013;10:1108–15.
noprecipitation experiments. Genomics 2004;83:349–60.
167. Wen PY, Lee EQ, Reardon DA, et al. Current clinical develop-
149. Ho JW, Bishop E, Karchenko PV, et al. ChIP-chip versus ChIP-
ment of PI3K pathway inhibitors in glioblastoma. Neuro
seq: lessons for experimental design and data analysis. BMC
Oncol 2012;14:819–29.
Genomics 2011;12:134.
168. Ostlund G, Sonnhammer EL. Avoiding pitfalls in gene (co)ex-
150. Bar-Joseph Z, Gerber GK, Lee TI, et al. Computational discov-
pression meta-analysis. Genomics 2014;103:21–30.
ery of gene modules and regulatory networks. Nat Biotechnol
169. Sun D, Luo M, Jeong M, et al. Epigenomic profiling of young
2003;21:1337–42.
and aged HSCs reveals concerted changes during aging that
151. Gao F, Foat BC, Bussemaker HJ. Defining transcriptional net-
reinforce self-renewal. Cell Stem Cell 2014;14:673–88.
works through integrative modeling of mRNA expression
170. Kowalczyk MS, Tirosh I, Heckl D, et al. Single-cell RNA-seq
and transcription factor binding data. BMC Bioinformatics
reveals changes in cell cycle and differentiation programs
2004;5:31.
upon aging of hematopoietic stem cells. Genome Res
152. Boulesteix AL, Strimmer K. Predicting transcription factor
activities from combined analysis of microarray and ChIP 2015;25:1860–72.
data: a partial least squares approach. Theor Biol Med Model 171. Deelen P, Zhernakova DV, de Haan M, et al. Calling geno-
2005;2:23. types from public RNA-sequencing data enables identifica-
153. Wu G, Ji H. ChIPXpress: using publicly available gene expres- tion of genetic variants that affect gene-expression levels.
sion data to improve ChIP-seq and ChIP-chip target gene Genome Med 2015;7:30.
ranking. BMC Bioinformatics 2013;14:188. 172. Navin NE. The first five years of single-cell cancer genomics
154. Karlebach G, Shamir R. Constructing logical models of gene and beyond. Genome Res 2015;25:1499–507.
regulatory networks by integrating transcription factor-DNA 173. Shapiro E, Biezuner T, Linnarsson S. Single-cell sequencing-
interactions with expression data: an entropy-based ap- based technologies will revolutionize whole-organism sci-
proach. J Comput Biol 2012;19:30–41. ence. Nat Rev Genet 2013;14:618–30.
155. Westra HJ, Peters MJ, Esko T, et al. Systematic identification 174. Stegle O, Teichmann SA, Marioni JC. Computational and
of trans eQTLs as putative drivers of known disease associ- analytical challenges in single-cell transcriptomics. Nat Rev
ations. Nat Genet 2013;45:1238. U1195. Genet 2015;16:133–45.
156. Consortium GT. Human genomics. The Genotype-Tissue 175. Teng M, Love MI, Davis CA, et al. A benchmark for RNA-seq
Expression (GTEx) pilot analysis: multitissue gene regula- quantification pipelines. Genome Biol 2016;17:74.
tion in humans. Science 2015;348:648–60. 176. Bernstein BE, Stamatoyannopoulos JA, Costello JF, et al. The
157. Alipanahi B, Delong A, Weirauch MT, et al. Predicting the se- NIH roadmap epigenomics mapping consortium. Nat
quence specificities of DNA- and RNA-binding proteins by Biotechnol 2010;28:1045–8.
deep learning. Nat Biotechnol 2015;33:831–8. 177. Bersanelli M, Mosca E, Remondini D, et al. Methods for the
158. Agarwal V, Bell GW, Nam JW, et al. Predicting effective integration of multi-omics data: mathematical aspects. BMC
microRNA target sites in mammalian mRNAs. Elife 2015;4: Bioinformatics 2016;17 (Suppl 2):15.

Co Expresion

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

Co Expresion

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Co Expresion

Uploaded by

Copyright:

Available Formats

Briefings in Bioinformatics, 19(4), 2018, 575–592

Gene co-expression analysis for functional

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Introduction sequencing (RNA-seq), and their respective data-analysis meth-

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

RNA-seq data for co-expression networks Clustering and network analysis

and mice to reveal a strong overlap between co-expression Guilt by association

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Differential co-expression analysis

Table 1. Methods and tools for RNA-seq-based co-expression network analysis

Tool/method Description, strengths (þ) and limitations ()

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

This is a non-comprehensive list of available tools and methods.

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Differential co-expression without prior grouping Comparison of differential co-expression

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

Downloaded from https://academic.oup.com/bib/article/19/4/575/2888441 by Universidad de Malaga user on 08 December 2021

You might also like