Báo cáo khoa học: TICL – a web tool for network-based interpretation of compound lists inferred by high-throughput metabolomics doc

The major advance of TICL is that it not only provides a model of possible compound transfor-mations related to the input list, but also implements a robust statistical framework to esti

TICL – a web tool for network-based interpretation of

compound lists inferred by high-throughput metabolomics Alexey V Antonov1, Sabine Dietmann1, Philip Wong1and Hans W Mewes1,2

1 Helmholtz Zentrum Mu¨nchen, Institute for Bioinformatics and Systems Biology, Neuherberg, Germany

2 Department of Genome-Oriented Bioinformatics, Technische Universita¨t Mu¨nchen, Freising, Germany

Knowledge of the molecular basis of metabolism is

crucial for our understanding of most cellular

pro-cesses [1–3] In recent years, technologies have been

developed that allow the systematic investigation of

large numbers of different metabolites [1,4–6] This has

led to metabolomics becoming an attractive technology

for exploring the molecular basis of complex cell

disor-ders [7–10]

In most genomics and proteomics studies aimed at

deciphering the molecular mechanisms of complex

bio-logical phenomena, the output is usually a list of

genes⁄ proteins [11–13] The next common step is the

application of bioinformatics and statistical methods

to obtain a statistically valid interpretation of the

derived gene list There are dozens of bioinformatics

tools available for the interpretation of gene lists A standard solution is the inference of over-⁄ under-repre-sented gene ontology terms [14–22] The significance of the produced results is usually supplied in the form of

a P-value The P-value represents a probability of inferring a similar or greater enrichment (for any gene ontology term) for a randomly sampled gene list [19] More complex methods have been proposed to exploit the database information currently available for metabolic and signaling pathways, such as the Kyoto Encyclopedia of Genes and Genomes (KEGG) [23] or BioCarta (http://www.biocarta.com) In this case, pathway topology was taken into account by developing specialized scoring functions The method developed by Rahnenfuhrer et al [24] includes, in

Keywords

bioinformatics tools for high-throughput

metabolomics; metabolomics; statistical

analysis and data mining; statistical and

bioinformatics tools; web tools for

metabolomics

Correspondence

A V Antonov, Helmholtz Zentrum Mu¨nchen

– German Research Center for

Environmental Health (GmbH), Institute for

Bioinformatics and Systems Biology,

Ingolsta¨dter Landstraße 1, D-85764

Neuherberg, Germany

Fax: +49 89 3187 3585

Tel: +49 89 3187 2788

E-mail: a.antonov@helmholtz-muenchen.de

(Received 12 November 2008, revised 28

January 2009, accepted 2 February 2009)

doi:10.1111/j.1742-4658.2009.06943.x

High-throughput metabolomics is a dynamically developing technology that enables the mass separation of complex mixtures at very high resolu-tion Metabolic profiling has begun to be widely used in clinical research to study the molecular mechanisms of complex cell disorders Similar to trans-criptomics, which is capable of detecting genes at differential states, meta-bolomics is able to deliver a list of compounds differentially present between explored cell physiological conditions The bioinformatics chal-lenge lies in a statistically valid interpretation of the functional context for identified sets of metabolites Here, we present TICL, a web tool for the automatic interpretation of lists of compounds The major advance of TICL is that it not only provides a model of possible compound transfor-mations related to the input list, but also implements a robust statistical framework to estimate the significance of the inferred model The TICL web tool is freely accessible at http://mips.helmholtz-muenchen.de/proj/ cmp

Abbreviations

KEGG, Kyoto Encyclopedia of Genes and Genomes; SHR, spontaneously hypertensive rat; WKY, Wistar Kyoto rat.

addition, the distance between genes within the

meta-bolic pathway The impact of a pair of genes is

weighted with respect to the distance between genes

within the metabolic pathway Another procedure

(impact analysis) proposed recently by Draghici et al

[25,26] goes beyond gene pairs and fully captures the

topology of signaling pathways by propagating

the perturbations measured at gene levels through the

entire pathway This technique can capture

informa-tion about the posiinforma-tion of the genes on the pathway,

because perturbation of the genes at the top of the

sig-naling cascade will propagate through the entire

path-way, unlike perturbation of the downstream genes

Metabolomics is a relatively new ‘omics’ technology

Experimental studies of complex cell disorders, which

employ high-throughput metabolomics as a basic

instrument, have just started to appear Several studies

of different diseases have demonstrated the successful

application of metabolomics in clinical research [7–9]

There is no doubt that the number of such clinical

studies will grow exponentially in the near future

Similar to transcriptomics and proteomics,

meta-bolomics allows for the detection of a list of markers,

present at different concentrations under various

explored cell physiological conditions In the case of

metabolomics, the markers are compounds (not genes

or gene products) There is a great demand for

bioin-formatics to provide a statistically valid interpretation

of compound lists produced experimentally Currently,

several bioinformatics approaches are available for

metabolomics Each approach was developed to solve

different practical problems related to the analysis of

metabolomics data [5,27–30] Most of the proposed

tools for metabolomics deal with the mass peak

anno-tation problem [31] The MassTrix web server has

recently been presented [30] and provides the

possibil-ity of uploading a high-precision mass spectrum,

auto-matically annotating mass peaks and mapping

identified compounds onto KEGG metabolic

path-ways Most of the available tools aim to interpret the

whole mass spectra rather than a sparse list of

com-pounds differentially present between samples Other

tools are available that provide visualizations of a

compound list in the context of metabolic networks

[32,33] The KEGG atlas accepts a list of compounds

as an input The output of the KEGG atlas is a

graph-ical visualization of compounds in the context of the

global metabolic reaction network The KEGG atlas,

however, does not provide quantitative and statistical

analyses

It is important to know whether experimentally

selected compounds are related, for example, whether

they belong to a chain or network of metabolic

reac-tions A partial answer to this question can be obtained from the KEGG atlas However, without quantitative analysis, there are no clues about the quality of these relations To fill the gap, we propose

an analytical framework for the interpretation of molecular mechanisms that unite a list of compounds This analytical framework is implemented as the freely accessible web tool TICL As we demonstrate using data from recently published metabolomics studies, TICL translates compounds into a set of linked meta-bolic reactions and provides quantitative estimates of the significance of the inferred models

Results

We consider several recently published experimental studies that report lists of compounds found to be dif-ferentially present under diverse physiological condi-tions We demonstrate that the proposed statistical framework can be helpful in understanding the biologi-cal context of the reported compound lists We start with the study by Lu et al [9], which reports metabolic variation related to hypertension and age-related conditions To characterize the development of hyper-tension, the spontaneously hypertensive rat (SHR), and its normotensive control, the Wistar Kyoto (WKY) rat, were investigated, and their blood plasma was analyzed using GC⁄ time-of-flight MS In total,

187 peaks were quantitatively determined after decon-volution, and 78 of them were identified Plasma com-positional differences for many identified compounds showed significant age-related variations for both SHR and WKY Also, many identified compounds showed significant variations between hypertension-related SHR and control WKY rats

Table 2 in Lu et al [9] reports 20 compounds that show significantly increased or decreased levels from 10

to 18 weeks of age in both SHR and WKY rats In total, 16 compounds can be mapped to the global com-pound network inferred from the KEGG Submission

of this list to the KEGG atlas gives the graphical visu-alization presented in Fig 1 At first glance, these com-pounds have nothing in common; they do not represent any specific canonical metabolic pathway In this case, visual analyses of Fig 1 cannot give a clear answer as

to whether and how the compounds are related By contrast, submission of this list to the TICL gives quantitative values that describe the quality of the rela-tions between the input compounds and provides a confidence score for such relations in the form of a P-value (the probability that randomly generated com-pound lists are involved in relations of similar quality) The report for the analyzed list is given in Table 1

Fig 1 Output returned by the KEGG atlas after submission of 20 compounds that have significantly increased or decreased levels from 10

to 18 weeks of age in both SHR and WKY rats Red points correspond to submitted compounds.

From Table 1 we can see the dependency between

the numbers of input compounds, which are involved

in the network model, and the number of allowed

missing compounds between any two input compounds

to be considered connected For example, we can

deduce that only two compounds (model 1) from the

input list are related as substrate and product of the

same reaction If one missing compound is allowed, a

maximum of four compounds from the input list are

connected into a network (model 2) For example,

model D5, which allows up to four intermediate

com-pounds, covers 11 metabolites For each model, the

P-value was estimated using a Monte Carlo procedure

For the most significant model D5, the estimated

P-value was < 0.01 This means that when we

ran-domly sampled a list of 16 compounds 100 times (only

compounds from the global compound network were

used to sample a random list) and applied the network

inference procedure to the random list, there was no

case, whereas the size of the inferred model D5from a

random list is 11 In all these cases, it was less Thus,

the P-value suggests that these 11 compounds

repre-sent a statistically valid metabolic network model

TICL provides a number of online visualization

capa-bilities The user can also download a preformatted

text file and use the medusa package [34] to visualize

the inferred model on a computer Figure 2 illustrates

a typical visualization output (model D5)

Table 3, in Lu et al [9], reports 22 compounds whose levels were significantly different between SHR and WKY rats In total, 14 compounds can be mapped to the global compound network inferred from the KEGG Submission of this list to the KEGG atlas gives the graphical visualization presented in Fig 3 Again, visualization of these compounds on the global metabolic network is not sufficient to obtain a full understanding of the quality of the relations among the compounds The report for the analyzed list

is presented in Table 2

From Table 2, we can see that the second set of compounds with significantly different levels between hypertensive (SHR) and control (WKY) rats does not define a statistically robust transformation network For example, model D6, which allows up to five miss-ing compounds between any two compounds from the input list, covers only eight input metabolites The statistical significance of the inferred models (for all models D1, , D6) was insignificant (the most signifi-cant model, D5, covers seven compounds, P > 0.1) The identified compounds are related to each other, although no more so than randomly selected com-pounds Thus, in the first case (age-related differences), TICL provides statistically valid arguments that the identified metabolites represent a set of dependent compounds Most probably, the identified compounds reflect structural, age-related changes in metabolism, in which whole metabolic blocks function differently In the second case (differences between SHR and WKY rats), however, no indication of structural metabolic variations can be found We admit that the result might have been influenced by the incomplete informa-tion currently available for metabolic reactions Another reason might be that the identified markers

do not necessarily reflect structural metabolic varia-tions, because there might be more complex mecha-nisms, not directly related to metabolism, which actually unite these compounds

The next example considered is related to a clinical study [7] In this study, a set of 66 invasive ovarian carcinomas and 9 borderline tumors of the ovary were analyzed by GC⁄ time-of-flight MS After automated mass spectral deconvolution, 291 metabolites were detected, of which 114 (39.1%) were annotated as known compounds Using a t-test, 51 metabolites were identified to be significantly (P < 0.01) different between borderline tumors and carcinomas Table 1, in Denkert et al [7], reports 26 significantly different metabolites which are known, 21 of which are mapped

to the global metabolic network The standard output

Table 2 The quantitative report ‘Enriched subnetworks’ returned

by TICL after the submission of 22 compounds with significantly

different levels between SHR and WKY rats.

Model

Maximum distance

between compounds

No input compounds

in the subnetwork P-value

Table 1 The quantitative report ‘Enriched subnetworks’ returned

by TICL after the submission of 20 compounds with significantly

increased or decreased levels from 10 to 18 weeks of age in both

SHR and WKY rats.

Model

Maximum distance

between compounds

No input compounds

in the subnetwork P-value

report from TICL for these compounds is given in

Table 3

If we consider the metabolite pathway membership,

then only ‘Nitrogen metabolism’ is presented in the list

more then twice Nevertheless, from Table 3 we can

see that almost all of the identified known metabolites

are dependent For example, model D2, which allows

only one missing metabolite, covers eight compounds

from the input list Model D3, which allows only two

missing metabolites, covers 15 input compounds and

model D4 covers almost all (19 of 21) metabolites

Figure 4 illustrates a typical visualization output for model D4

The last example we consider is related to another clinical cancer study In this case, the target was colon carcinoma A set of paired samples of normal colon and colorectal cancer tissue was investigated by

GC⁄ time-of-flight MS, which allowed robust detection

of a total of 206 metabolites Subsequent analysis revealed that 82 metabolites were significantly different Table 4 presents TICL output for these 82 compounds We can see that almost all of the identi-fied known metabolites are dependent For example, model D2, which allows only one missing metabolite, covers 37 compounds from the input list Model D3, which allows only two missing metabolites, covers 49 input compounds Figure 5 illustrates a typical visuali-zation output produced using TICL for model D3

In both cancer-related examples, TICL provides statistically valid arguments that the identified meta-bolites represent a set of dependent compounds Although the analyzed cases were related to different tissues (ovarian cancer and colon cancer), in both cases, the discovered metabolic markers were not inde-pendent; they define a related set of metabolic reac-tions which, in turn, define a semi-noninterrupted

Fig 2 Visualization of the inferred network model D 5 returned by TICL after submission

of 20 compounds that have significantly increased or decreased levels from 10 to

18 weeks of age in both SHR and WKY rats Boxes are compounds from the input list, circles are intermediate compounds Colors are used to specify canonical KEGG meta-bolic pathways.

Table 3 The quantitative report ‘Enriched subnetworks’ returned

by TICL on submission of 21 known compounds found to have

sig-nificantly different concentrations between borderline ovarian

tumors and ovarian carcinomas.

Model

Maximum distance

between compounds

No input compounds

in the subnetwork P-value

Fig 3 The output returned by KEGG atlas after submission of 22 compounds that have levels significantly different between SHR and WKY rats Red points correspond to the submitted compounds.

network of metabolic transformations that covers most

of the identified compounds Thus, in these two cases,

TICL provides new biological insights into variations

in metabolic processes in cancer and presents statistical arguments validating these insights

Discussion

In addition to the ability to generate a large amount

of data per experiment, high-throughput technologies also brought the challenge of translating such data into a better understanding of the underlying biologi-cal phenomena A number of tools in the field of transcriptomics and proteomics have been developed recently to interpret gene⁄ protein lists in order to address this challenge High-throughput metabolomics has recently started to be instrumental in exploring metabolic variations on a genomic scale [7–10,35,36] The output produced by experimental metabolomics is similar to other ‘omics’ technologies in the sense that

Fig 4 Visualization of the inferred network model D4returned by TICL after submission of 21 compounds found to have significantly different concentrations in borderline ovarian tumors and carcinomas Boxes are compounds from the input list, circles are intermediate compounds Colors are used to specify canonical KEGG metabolic pathways.

Table 4 The quantitative report ‘Enriched subnetworks’ returned

by TICL on submission of 82 known compounds found to have

significantly different concentrations between normal colon tissue

and colorectal cancer tissue.

Model

Maximum distance

between compounds

No input compounds

in the subnetwork P-value

it provides a list The difference is that it is not a

gene⁄ protein list, but a list of compounds, whose

con-centration differs between the considered cell (tissue)

phenotypes

The bioinformatics tools and procedures currently

available in the field of metabolomics are more

rele-vant for the annotation of mass peaks or for the

inter-pretation of whole mass peaks spectra To our

knowledge, there is currently no procedure or tool

available that deals with a relatively sparse compound

list found to be differentially present between different

cell physiological conditions As demonstrated here,

such lists can be translated into network models, which

cover most metabolites from the supplied list How-ever, the sparseness of the compound list presumes that the inferred models may have a lot of intermedi-ate compounds (up to 2–5 intermediintermedi-ate compounds between any two compounds from the input list cov-ered by the model) In this case, tools that offer only a visualization of compounds in the context of the global metabolic network are inefficient It is evident that if relaxing the number of possible missing compounds, sooner or later, one will be able to cover all input compounds It is essential to provide a model of the possible metabolic transformations that cover the input compound list, and also to estimate quantitatively the

Fig 5 Visualization of the inferred network model D 3 returned by TICL after submission of 82 compounds found to have significantly differ-ent concdiffer-entrations in normal colon tissue and colorectal cancer tissue Boxes are compounds from the input list, circles are intermediate compounds Colors are used to specify canonical KEGG metabolic pathways.

quality of the produced model TICL is the first tool

for the analysis of compound lists that implements

such quality control by providing P-values for the

inferred models

Materials and methods

Given a compound list found to be differentially present

between biological samples, we translate this list into a

network model In other words, we reconstruct the most

probable transformation routes that unite compounds from

the list In some sense, this task is similar to the problem of

finding the shortest path between two compounds, but is

extended to list of compounds [27,37] To restore the

trans-formation routes, we use a global metabolic network

inferred from the KEGG database The major advance of

TICL is that it not only provides a model of possible

com-pound transformations related to the input list, but also

implements a robust statistical framework to estimate the

significance of the inferred model In simple terms, the

P-values inferred by Monte Carlo simulations [17,38,39]

represent the probability of a random list having the same

quality model

Global compound network

The KEGG REACTION database is a collection of

chemi-cal structure transformation patterns for substrate–product

pairs (reactant pairs) We can build a global ‘reaction

network’ (reactions are nodes, compounds are edges) by

con-necting edges and reactions that share the same compounds

In general, a reaction consists of multiple reactant pairs, and

the one that appears in a KEGG metabolic pathway is called

a main pair To build a global reaction network, we used

only compounds classified as main reaction pairs

Network inference procedure

At the start of the procedure, we have a list of compounds

(the input list), on the one hand, and the global compound

network, on the other hand The distance between two

arbitrary compounds is computed as the minimum number

of consecutive steps required to get from one compound to

another by working through existing paths on the global

compound network Distance 1 means that the two

com-pounds are directly connected (related as substrate and

product of a metabolic reaction); distance 2 means that the

two compounds are connected via one intermediate

com-pound; distance 3 means that the two compounds are

con-nected via two intermediate compounds, and so on Given

a compound list, our purpose is to infer the network model

(connect some pairs from the input list to get connected

component) that minimizes the distance between each

connected pair of compounds

Initially, we map compounds from the input list onto the global compounds network At this point, all compounds from the input list are disconnected In the first step, all pairs of compounds with distance 1 are connected by edges and we look for connected subnetworks The subnetwork with the maximal number of compounds is referred to as

an inferred network model D1 In the second step, com-pounds (from the input list) with distance 2 are connected

by edges The subnetwork with the maximal number of compounds is inferred and referred to as network model

D2 In a similar way, network models D3, D4, up to a spec-ified number z (model Dz) are inferred Models D2, D3, ,

Dz incorporate compounds that are not from the input list but are added to connect input compounds in the network model We refer to these added compounds as intermediate

or missing compounds

Statistical treatment

Let us assume that we have an input compound list of size N and using the network inference procedure described above we infer the network models D1, , Dz, which allow

0, 1, , z - 1 intermediate compounds to be added to the model Let us denote S1, S2, , Sz to be the number of input compounds in the inferred network models We also refer to S1, S2, , Sz as the sizes of the respective models

D1, , Dz Given the size of the input compound list (N),

we consider the sizes of the models (values S1, S2, , Sz) to

be quality measures We have to estimate the probability of inferring models of the same or larger sizes from randomly generated compound lists of size N

To estimate the significance of the inferred models, we compare the values S1, S2, , Szwith background distribu-tions BD1, , BDzcomputed using Monte Carlo simulation [39] To generate the background distributions BD1, ,

BDz, we repeat the following simulation procedure k times, where k specifies the upper significance level A random gene list Ljof size N (equal to the size of the input list) is generated by sampling compounds from global compound network Index j = 1 k specifies each of the k random simulations The network inference procedure described above is applied to the random list Lj and the network models D1, , Dzare inferred Let us denote the size (the number of input compounds) of the inferred models D1, ,

Dzfor the random list Ljas R1j, ., Rzj Thus, after repeat-ing the simulation procedure k times, we get the background distribution R1j(j = 1 k) for models D1, the background distribution R2j(j = 1 k) for models D2, and the back-ground distribution Rzj(j = 1 k) for models Dz

To estimate significance of the inferred network model

D1 for the input gene list, the value S1 is compared with the distribution R1j Let n be the number of values from the distribution R1jthat are‡ S1.The estimate of P of the inferred network model D1is computed as P = (n + 1)⁄ k

In the same way, the P-values for models D2, , Dz are

computed using values S2, ,Sz and background

distribu-tions R2j , Rzj In other words, the P-value is estimated

as the share of random simulations where the size of the

inferred models for random compound lists of size N are

equal to or greater than the size S1, S2, , Sz of the

inferred models for input compound list (size N)

References

1 Fiehn O (2001) Combining genomics, metabolome

analysis, and biochemical modelling to understand

metabolic networks Comp Funct Genomics 2, 155–168

2 Goodacre R (2005) Metabolomics shows the way to

new discoveries Genome Biol 6, 354

3 Hertkorn N, Ruecker C, Meringer M, Gugisch R,

Frommberger M, Perdue EM, Witt M &

Schmitt-Kopplin P (2007) High-precision frequency measurements:

indispensable tools at the core of the molecular-level

analysis of complex systems Anal Bioanal Chem 389,

1311–1327

4 Fiehn O (2008) Extending the breadth of metabolite

profiling by gas chromatography coupled to mass

spec-trometry Trends Anal Chem 27, 261–269

5 Shulaev V (2006) Metabolomics technology and

bioin-formatics Brief Bioinform 7, 128–139

6 Shulaev V & Oliver DJ (2006) Metabolic and proteomic

markers for oxidative stress New tools for reactive

oxygen species research Plant Physiol 141, 367–

372

7 Denkert C, Budczies J, Kind T, Weichert W, Tablack

P, Sehouli J, Niesporek S, Konsgen D, Dietel M &

Fiehn O (2006) Mass spectrometry-based metabolic

profiling reveals different metabolite patterns in invasive

ovarian carcinomas and ovarian borderline tumors

Cancer Res 66, 10795–10804

8 Denkert C, Budczies J, Weichert W, Wohlgemuth G,

Scholz M, Kind T, Niesporek S, Noske A, Buckendahl

A, Dietel M et al (2008) Metabolite profiling of human

colon carcinoma – deregulation of TCA cycle and

amino acid turnover Mol Cancer 7, 72

9 Lu Y, Jiye A, Wang G, Hao H, Huang Q, Yan B, Zha

W, Gu S, Ren H, Zhang Y et al (2008) Gas

chromatography⁄ time-of-flight mass spectrometry based

metabonomic approach to differentiating

hypertension-and age-related metabolic variation in spontaneously

hypertensive rats Rapid Comm Mass Spectrom 22,

2882–2888

10 Altmaier E, Ramsay SL, Graber A, Mewes HW,

Wein-berger KM & Suhre K (2008) Bioinformatics analysis

of targeted metabolomics – uncovering old and new

tales of diabetic mice under medication Endocrinology

149, 3478–3489

11 Shi Q, Bao S, Song L, Wu Q, Bigner DD, Hjelmeland

AB & Rich JN (2007) Targeting SPARC expression

decreases glioma cellular survival and invasion associ-ated with reduced activities of FAK and ILK kinases Oncogene 26, 4084–4094

12 Marquez RT, Baggerly KA, Patterson AP, Liu J, Broaddus R, Frumovitz M, Atkinson EN, Smith DI, Hartmann L, Fishman D et al (2005) Patterns of gene expression in different histotypes of epithelial ovarian cancer correlate with those in normal fallopian tube, endometrium, and colon Clin Cancer Res 11, 6116– 6126

13 LaTulippe E, Satagopan J, Smith A, Scher H, Scardino

P, Reuter V & Gerald WL (2002) Comprehensive gene expression analysis of prostate cancer reveals distinct transcriptional programs associated with metastatic dis-ease Cancer Res 62, 4499–4506

14 Adler P, Reimand J, Janes J, Kolde R, Peterson H

& Vilo J (2008) KEGGanim: pathway animations for high-throughput data Bioinformatics 24, 588–590

15 Antonov AV & Mewes HW (2006) Complex functional-ity of gene groups identified from high-throughput data

J Mol Biol 363, 289–296

16 Antonov AV, Schmidt T, Wang Y & Mewes HW (2008) ProfCom: a web tool for profiling the complex functionality of gene groups identified from high-throughput data Nucleic Acids Res 36, W347–W351, doi:10.1093/nar/gkn239

17 Antonov AV & Mewes HW (2008) Complex phyloge-netic profiling reveals fundamental genotype–phenotype associations Comput Biol Chem 32, 412–416

18 Khatri P, Draghici S, Ostermeier GC & Krawetz SA (2002) Profiling gene expression using onto-express Genomics 79, 266–270

19 Khatri P & Draghici S (2005) Ontological analysis of gene expression data: current tools, limitations, and open problems Bioinformatics 21, 3587–3595

20 Khatri P, Voichita C, Kattan K, Ansari N, Khatri A, Georgescu C, Tarca AL & Draghici S (2007) Onto-Tools: new additions and improvements in 2006 Nucleic Acids Res 35, W206–W211

21 Reimand J, Kull M, Peterson H, Hansen J & Vilo J (2007) g:Profiler – a web-based toolset for functional profiling of gene lists from large-scale experiments Nucleic Acids Res 35, W193–W200

22 Reimand J, Tooming L, Peterson H, Adler P & Vilo J (2008) GraphWeb: mining heterogeneous biological net-works for gene modules with functional significance Nucleic Acids Res 36, W452–W459, doi:10.1093/nar/ gkn230

23 Ogata H, Goto S, Sato K, Fujibuchi W, Bono H & Kanehisa M (1999) KEGG: Kyoto encyclopedia of genes and genomes Nucleic Acids Res 27, 29–34

24 Rahnenfuhrer J, Domingues FS, Maydt J & Lengauer

T (2004) Calculating the statistical significance of changes in pathway activity from gene expression data Stat Appl Genet Mol Biol 3, Article 16