Network analysis identifies a putative role for the PPAR and type 1

interferon pathways in glucocorticoid actions in asthmatics

Supplementary material

Diego Diez1†, Susumu Goto1, John V. Fahy2,4, David J. Erle2,3,4, Prescott G. Woodruff2,4, Åsa M. Wheelock5*, Craig E. Wheelock6*

1 Bioinformatics Center, Institute for Chemical Research, Kyoto University,

Uji, Kyoto, 611-0011, Japan 2 Division of Pulmonary and Critical Care Medicine, Department of

Medicine, University of California San Francisco, USA 3 Lung Biology Center, University of California San Francisco, USA 4 Cardiovascular Research Institute, University of California San Francisco,

USA 5 Respiratory Medicine Unit, Department of Medicine, Karolinska Institutet,

Stockholm, Sweden 6 Department of Medical Biochemistry and Biophysics, Division of

Physiological Chemistry II, Karolinska Institutet, Stockholm, Sweden

†Current address: Laboratory of Bioinformatics and Genomics, World Premier International Immunology Frontier Research Center, Osaka University, Osaka 565-0871, Japan

* Corresponding author: Craig E. Wheelock: craig.wheelock@ki.se

Åsa M. Wheelock: asa.wheelock@ki.se

Construction of the co-transcriptional network

A co-transcriptional network was constructed from the microarray data using the ARACNE method, which uses mutual information (MI) to test statistical dependency between expression profiles [2]. Mutual information was computed using a Gaussian Kernel estimate. Two critical parameters for the ARACNE method are the kernel width and the MI threshold, which is the minimum MI value to be considered statistically significant. Both parameters are normally estimated using precomputed calibration curves that were generated for a different microarray platform. To avoid bias due to the platform differences and ensure the best parameters for our dataset are used, custom calibration curves were constructed using the MATLAB scripts provided in the original publication [2]. A calibration curve for kernel width was generated (Figure S11A) and the optimal kernel width was determined to be 0.2179931, which is 5.3% lower than that estimated by ARACNE (0.230167). The calibration curve for the MI threshold was computed (Figure S11B) and used during network reconstruction. Next, the optimal p-value cutoff for network reconstruction was determined by computing networks at different p-values over a sufficiently wide range. For each network, the number of nodes, edges and degree distribution was computed, and a p-value of 10-8 was chosen, which resulted in a power law degree distribution for the network (Figure S12). Finally, in order to assess the effect of sample ordering upon correlation estimations (MI), a bootstrapped network was computed by setting 100 cycles and the consensus network was calculated with a Bonferroni corrected q < 0.001 [3]. The inferred network contained 7,676 nodes and 49,499 edges (Figure S1), and the degree distribution followed a power-law with a gamma parameter equal to 1.69 (Figure S2).

Module detection in the co-transcriptional network

Clustering of the co-transcriptional network was performed using MCODE. Several modules were selected based on statistical, functional and relevance to asthma, inflammation, epithelium and metabolism of steroids. Figure S1 shows the location of some MCODE modules (M1, M5, M6, M9) and the BioNet module in the co-transcriptional network, highlighting the dependency between modules M1 and M6. Figure S4 shows the expression profiles of modules M1, M5 (for up and down-regulated genes) and M6, indicating that M1 and M6 genes have highly correlated expression profiles. Figures S3, S5 and S6 show modules M5, M6 and M9 with the selected KEGG pathways and top-ten ranked GO biological processes detected by enrichment analysis.

Filtering of non-expressed genes

Non-expressed genes were eliminated on the basis of XIST expression, which is not expressed in males [1]. The distribution of intensities for XIST in male and females was computed (Figure S10), and the background threshold estimated as the median value of the distribution in males (5.32). Genes were filtered when the intensity in all arrays was below that of the determined threshold value.

Expression profiles for selected genes

The most relevant pathways selected in our analysis were Toll-like receptor signaling pathway and PPAR signaling. Four genes associated with these pathways were identified in the BioNet module. Figure S8 show the expression profiles of STAT1 and IRF9, which mediate signaling by TLR7 and TLR4, and are involved in the production of type 1 interferons. Both genes have correlated profiles that are also somewhat correlated with treatment. Figure S9 shows RXRA and PPARGC1A, which mediate signaling by PPARs, and are involved in the inhibition of inflammatory responses. Both profiles are correlated and correlated as well with treatment.

Figure S1. Summary of functional enrichment analysis. Heatmap displaying the gene count distribution of KEGG pathways enriched in the gene lists from the different analyses. RAW = the list of genes based upon P-values that were not corrected for multiple testing (P<0.05). FDR = the list of genes following multiple testing correction (q<0.05). BIONET = the module resulting from the BioNet analysis of the PPI network. M1 = module that had the highest score following MCODE analysis of the co-transcriptional network.

Endoc

ytosis

Antige

n proc

essin

g and

pres

entat

ion

Allogra

ft reje

ction

Glycoly

sis / G

lucon

eoge

nesis

Glutath

ione m

etabo

lism

Arginin

e and

proli

ne m

etabo

lism

Aldoste

rone-r

egula

ted so

dium re

abso

rption

Protea

some

Sulfur

metabo

lism

Keratan

sulfa

te bio

synth

esis

Adipoc

ytokin

e sign

aling

pathw

ay

ErbB si

gnali

ng pa

thway

Steroid

horm

one b

iosyn

thesis

Primary

bile

acid

biosy

nthes

is

Retino

l meta

bolism

Comple

ment a

nd co

agula

tion c

asca

des

Cell ad

hesio

n mole

cules

(CAMs)

Cytokin

e-cyto

kine r

ecep

tor in

terac

tion

PPAR sign

aling

pathw

ay

Toll-lik

e rec

eptor

sign

aling

pathw

ay

Chemok

ine si

gnali

ng pa

thway

Lyso

some

Intes

tinal

immun

e netw

ork fo

r IgA pr

oduc

tion

Leuk

ocyte

trans

endo

thelia

l migr

ation

RAWFDR

BIONET M1

0 5 10 15 20 25# Gene count

Figure S2. Co-transcriptional network. ARACNE-based network with selected modules visualized. Color key: M1 (yellow), M5 (cyan), M6 (green), M9 (red) and BioNet (pink). The functional relation between MCODE modules M1 and M6 is highlighted by the close proximity of the modules in the whole co-transcriptional network.

Figure S3. Co-transcriptional network degree distribution. The degree distribution for the ARACNE inferred co-transcriptional network follows a power law with gamma = -1.69.

Figure S4. MCODE module M5. All genes in this module are differentially expressed with q<0.05. Therefore, this module was discarded from the common pathways approach. Enrichment analysis returned ‘Steroid hormone biosynthesis’ as the most significantly enriched pathway, overlapping with the most significant pathway enriched in the FDR list.

1 2 5 10 20 50 200

1

510

50100

5001000

Degree

Count

R2 = 0.93, y = 3.77 − 1.69x

OSTalpha

EPDR1

KLF9

PKP2

SRD5A1

TMPRSS4GSDMC

SLC19A2

CYP4B1

FAM107A

ARXCCDC80

ENPP5

UBDMUC13

AHNAK2FXYD1

ERBB4

NFE2L1

GXYLT2

C10orf26

SMAP2

PDK4PHACTR3

SULT2B1

TFCP2L1

KEGGSteroid hormone biosynthesis

GeneOntologycerebral cortex GABAergic interneuron migrationcytoplasm organizationembryonic olfactory bulb interneuron precursor migrationepithelial cell maturationglobus pallidus developmentpositive regulation of synaptic transmission, GABAergicsteroid metabolic processsubstrate!independent telencephalic tangential migrationsynapse maturationthiamin transport

Figure S5. Expression profiles of selected MCODE modules. Profile plots of genes in modules M1 (darkgreen), M6 (black) and BioNet (darkred: up-regulated in flovent; darkblue: down-regulated in flovent). Functional relation between M1 and M6 is evidenced by they correlated profiles. Additionally, their independence from drug variable is also revealed. The genes in module M5, on the other hand, show correlation with treatment.

Figure S6. MCODE module M6.

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

−3−1

13

Profiles

M1M6M5/upM5/down

IFIT2

CLEC12A

SASH3

NCF2

IGSF6

HLA!DRANCKAP1L

GLDN

FLNA

CFD

RBP4

ADCY7

LHFPL2

SH2B3

PIK3AP1

CCRL2

IL12RB1

EMILIN2

LIPA

C7orf58

GGTA1

AOC3 RAC2

GIMAP4

TUBB6

ITGA5

BTKRBP7

UPP1

GPR137BPPARG

SLC31A2GM2A

STAC

MLKLFCGR3B

NFAM1

TLR4

NCF4

KEGG B cell receptor signaling pathwayFc epsilon RI signaling pathwayLeukocyte transendothelial migration

GeneOntologyactivation of immune responsecytokine productionpeptidyl!tyrosine phosphorylationpositive regulation of bindingpositive regulation of nucleocytoplasmic transportpositive regulation of protein transportpositive regulation of transcription factor activityregulation of DNA bindingregulation of transcription regulator activityresponse to external stimulus

Figure S7. MCODE module M9.

Figure S8. RXRA and PPARGC1A expression profiles. RXRA and PPARGC1A show correlated expression profiles and are upregulated in flovent.

GeneOntologypositive regulation of nitrogen compound metabolic processpositive regulation of RNA metabolic processpositive regulation of transcriptionresponse to cAMPresponse to corticosterone stimulusresponse to cytokine stimulusresponse to mechanical stimulusresponse to organic cyclic substanceresponse to progesterone stimulusSMAD protein signal transductionKEGG B cell receptor signaling pathwayMAPK signaling pathwayT cell receptor signaling pathwayToll!like receptor signaling pathway

IER3

FOS

EGR1

IER2

BTG2 JUNB

JUN

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Flovent

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

Placebo

−2−1

01

2

Profiles

RXRAPPARGC1A

Figure S9. Filtering of non-expressed genes was performed based on the expression of the gene XIST, which is missing in males and provides a lower intensity threshold.

Figure S10. A) Calibration curve for the kernel width computed for different samples sizes. B) Calibration curve for optimal MI threshold computation. Each measurement is performed in triplicate at each sample size.

5 6 7 8 9

0.0

0.5

1.0

1.5

2.0

XIST expression

N = 58 Bandwidth = 0.06292

Den

sity

MenWomen

0 5 10 15 20 25 30

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Sample Size

Kern

el W

idth

Drug (R2 = 0.78, y = 0.26 0.0013x)

10 15 20 25 30

-45

-40

-35

-30

-25

-20

Sample size

Slop

e

Gradient (R2 = 0.99, y = -12.68 1.09x)BA

Figure S11. A) Number of nodes and edges vs. P value profile for the drug co-transcriptional network. B) Degree distribution for networks constructed at different P value cutoffs. These plots show that at high P values the networks behave as random networks, with a number of nodes equal to the total number of genes in the array, and a large number of edges. As the P value decreases, the networks reduce the number of edges, and eventually the number of nodes (left). At the same time, the behavior of the networks changes from random to a scale free (right; blue networks). Finally, the number of nodes and edges is too low, and the network looses its structures.

References

1. Brown CJ, Hendrich BD, Rupert JL, Lafreniere RG, Xing Y, Lawrence J, Willard HF: The human XIST gene: analysis of a 17 kb inactive X-specific RNA that contains conserved repeats and is highly localized within the nucleus. Cell 1992, 71:527-542.

2. Margolin AA, Nemenman I, Basso K, Wiggins C, Stolovitzky G, Dalla Favera R, Califano A: ARACNE: an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context. BMC Bioinformatics 2006, 7 Suppl 1:S7.

3. Bonferroni CE: Il calcolo delle assicurazioni su gruppi di teste. Studi in Onore del Professore Salvatore Ortu Carboni 1935:13-60.

P va

lue

1e-05 1e-06 1e-07 1e-08 1e-09 1e-10 1e-11 1e-12

1e-12 1e-10 1e-08 1e-06 1e-041e+00

1e+02

1e+04

1e+06

P value

Cou

nts

EdgesNodes

P va

lue

1e-05 1e-06 1e-07 1e-08 1e-09 1e-10 1e-11 1e-12

1 5 50 500 50001

510

50100

5001000

5000

Degree

Cou

nt

A B