Andrey Alexeyenko

Gene interaction networks for functional analysis and

prognostication

“Data is not information, information is not knowledge,knowledge is not wisdom, wisdom is not truth,”

—Robert Royar (1994)

http://www.hdpaperz.com

Biological data production

…and analysis

FunCoup is a data integration framework to discover

functional coupling in eukaryotic proteomes with

data from model organisms

Amouse

Bmouse

?Fi

nd o

rthol

ogs*

Human

Fly

Rat

Yeast

Hig

h-th

roug

hput

ev

iden

ce

Andrey Alexeyenko and Erik L.L. Sonnhammer (2009) Global networks of functional coupling in eukaryotes from comprehensive data integration. Genome Research.

FunCoup• Each piece of data is evaluated• Data FROM many eukaryotes (7)• Practical maximum of data sources (>50)• Predicted networks FOR a number of eukaryotes

(10…)• Organism-specific efficient and robust Bayesian

frameworks• Orthology-based information transfer and

phylogenetic profiling• Networks predicted for different types of

functional coupling (metabolic, signaling etc.)

FunCoup: on-line interactome resource

Andrey Alexeyenko and Erik L.L. Sonnhammer (2009) Global networks of functional coupling in eukaryotes from comprehensive data integration. Genome Research.

http:

//Fu

nCou

p.sb

c.su

.se

Do gene networks tell

any story?

• Yellow diamonds: somatic mutations in prostate cancer

• Pink crosses: also mutated in glioblastome (TCGA)

State-of-the-art method to beat:Frequency analysis of somatic mutations

Network enrichment analysis:

Altered genes (green) from one individual lung cancer are enriched in network connections to members of ErbB (HER2) pathway (yellow) and GO term “apoptosis” (blue).

Question: How to find something in common between many altered genes?

Apophenia: the human propensity to see meaningful patterns in random data

(Brugger 2001; Fyfe et al. 2008)

Looks “interesting”…

Statistically significant (individually)

Statistically significant (adjusted for multiple testing)

Functionally relevant: THE CORE

Confidence of biological observations:

PLoS Med 2005 2(8):e124

• Dimensionality curse: many variables and few observations, while causative events are rare.

• Tradition not to publish negative results.• Acceptance of a common p-value threshold

without adjustment for multiple testing (a hundred papers with single experiment in each is multiple testing!..).

Remarkable failures of biological researchProblem to solve Reason for failure Proposed solutionPlant breeding for complex sets of advantageous traits

Extreme combinatorial complexity Set up much larger trials

Genome sequence -> function Lack of observed variation, perturbations Observe more sequence variability (novel genomes, SNPs, pedigrees etc.)

Sequence-> protein structure Computationally unfeasible ?

Reverse engineering of regulatory networks from a single data platform (gene expression)

There are many parallel mechanisms of regulation Invent more platforms, combine

Quantitative modeling in regulatory networks

Extreme topological complexity (most of ingoing nodes are not known)

Find more complete networks

Many constraints are structural, undirected Collect extra data: chromatin structure, histone modifications

Dependence on context (environment, cell milieu) Collect more conditions

Finding the “ultimate”, full cellular network

The “functional coupling” is loosely defined Collect whatever possible

Gene expression signatures predictive of e.g. disease outcome

Dimensionality curse Increase sample size

Genetic determinants of diseases Most of the genetics is in gene interactions, not possible to address with common “sample size” approaches

Increase sample size

From Speliotes et al. 2010: Genome-wide association results for the body mass index, meta-analysis(a) Manhattan plot showing the significance of association between all SNPs and BMI in the stage 1 meta-analysis, highlighting SNPs previously reported to show genome-wide significant association with BMI (blue), weight or waist circumference (green), and the 18 new regions described here (red). The 19 SNPs that reached genome-wide significance at Stage 1 (13 previously reported and 6 new) are listed in Table 1). (b) Quantile-quantile (Q-Q) plot of SNPs in stage 1 meta-analysis (black) and after removing any SNPs within 1 Mb of the 10 previously reported genome-wide significant hits for BMI (blue), after additionally excluding SNPs from the four loci for waist/weight (green) and after excluding SNPs from all 32 confirmed loci (red). The plot was abridged at the Y-axis (at P < 10−20) to better visualise the excess of small P-values after excluding the 32 confirmed loci (Supplementary Fig. 3 shows full-scale Q-Q plot). The shaded region is the 95% concentration band. (c) Plot of effect size (in inverse normally transformed units (invBMI)) versus effect allele frequency of newly identified and previously identified BMI variants after stage 1 + stage 2 analysis; including the 10 previously identified BMI loci (blue), the four previously identified waist and weight loci (green) and the 18 newly identified BMI loci (blue). The dotted lines represent the minimum effect sizes that could be identified for a given effect-allele frequency with 80% (upper line), 50% (middle line), and 10% (lower line) power, assuming a sample size of 123,000 individuals and a α-level of 5×10−8.

Even a cohort of 250.000 individuals with 2.800.000 SNPs could only explain less than 4% of variability of the body mass index (the total heritability had been estimated as 40-70%).The authors extrapolate that 730.000 individuals would explain around 5% of the observed variance).

…………………………………………………………………………………………………………………………

Yuri Lazebnik, 2002

How does it still work?

• “The hyperbrain”. Scientific community critically evaluates each finding, most commonly rejects it…

• In reality, scientists do not what they promise in grant applications.

• Scientific data is used not what it was meant for. • Common sense still rules.• Answering simple questions does the job.

Network enrichment analysis:

Altered genes (green) from one individual lung cancer are enriched in network connections to members of ErbB (HER2) pathway (yellow) and GO term “apoptosis” (blue).

Question: How to find something in common between many altered genes?

Network enrichment analysis: compared to a reference and quantified

N links_real = 12

N links_expected = 4.65

Standard deviation = 1.84

Z = (N links_observed – N links_expected) / SD = 3.98

P-value = 0.0000344

FDR < 0.1

Actual network: observed pattern A random pattern

Question:Is ANXA2 related to TGFbeta signaling?

How informative is a global gene network?

Functional characterization of novel gene sets

Functional set?

Our alternative:Network enrichment analysis

Altered genes

0 50 100 150 200 250 300 350 400 450

No. of positives, random groups

0

200

400

600

800

1000

1200

1400

No.

of p

ositi

ves,

rea

l gro

ups

GEA_SIGN GEA_REST NEA_SIGN NEA_REST

State-of-the-art method to beat:Gene set enrichment analysis

Alexeyenko A, Lee W, … Pawitan Y. Network enrichment analysis: extension of gene-set enrichment analysis to gene networks. BMC Bioinformatics, 2012

Primary molecular processes

Response, recorded with high-throughput

paltforms

Known biological units: processes, pathways

Protein abundance1. 2. 3. 4. …N.

Phosphorylation1. 2. 3. 4. …N.

Methylation1. 2. 3. 4. …N.

mRNA expression1. 2. 3. 4. …N.

ChipSeq1. 2. 3. 4. …N.

Mutation profile1. 2. 3. 4. …N.Metabolites

1. 2. 3. …N.

Pathway analysis: why needed, what it is?

Towards even better network predictionpartial correlations:

a way to get rid of spurious links

0.7

0.6

0.4

Cancer-specific networks:links inferred from

expression, methylation, mutations

Functional couplingtranscription transcription transcription methylation methylation methylation mutation methylation mutation transcriptionmutation mutation+ mutated gene

State-of-the-art method to beat:

Reverse engeneering froma single

source (usually transcriptome)

Now: answer biological questions

Molecular phenotypes in network space(lung cancer, data from ChemoRes consortium)

P53 signaling

Cell cycle

Apoptosis0 2 4 6 8

0.0

0.2

0.4

0.6

0.8

1.0

Years since surgery

Rel

apse

-free

sur

viva

l

Predictors:

GO:0001666 response to hypoxiaGO:0005154 EGFR bindingGO:0005164 TNF bindingGO:0070848 response to growth factor stimulus

Question: How to distinguish between different molecular subtypes of cancer?

Is NTRK1 a driver in the GBM tumor TCGA-02-0014?

Somatic mutations: drivers vs. passengersdata from The Cancer Genome Atlas

Gains and losses on chromosome 7, in 142 glioblastoma multiforme genomes, TCGA

Driver copy number changesCopy number of MAP3K11

Altered Normal

Point mutation inPTEN

Present 6 (~2) 29 (~33)Absent 2 (~6) 105 (~102)

Cop

y nu

mbe

rC

onfid

ence

Genetic association of sequence variants near AGER/NOTCH4 and dementia.Bennet AM, Reynolds CA, Eriksson UK, Hong MG, Blennow K, Gatz M, Alexeyenko A, Pedersen NL, Prince JA.J Alzheimers Dis. 2011;24(3):475-84.

Genome-wide pathway analysis implicates intracellular transmembrane protein transport in Alzheimer disease.Hong MG, Alexeyenko A, Lambert JC, Amouyel P, Prince JA.J Hum Genet. 2010 Oct;55(10):707-9. Epub 2010 Jul 29.

Analysis of lipid pathway genes indicates association of sequence variation near SREBF1/TOM1L2/ATPAF2 with dementia risk.Reynolds CA, Hong MG, Eriksson UK, Blennow K, Wiklund F, Johansson B, Malmberg B, Berg S, Alexeyenko A, Grönberg H, Gatz M, Pedersen NL, Prince JA.Hum Mol Genet. 2010 May 15;19(10):2068-78. Epub 2010 Feb 18.

Validation of candidate disease genes(work with Jonathan Prince, MEB, KI)

Question: Is there extra evidence for GWAS-candidates to be involved?Answer: Yes, for some…

Relapse

+ -

Treatment

+ A B

- C D

Resistance to vinorelbine in lung cancerT -R - T -R + T +R - T +R +

-4-3

-2-1

01

23

MED

17

T -R - T -R + T +R - T +R +

-6-5

-4-3

-2-1

0

AC

005921.3

T -R - T -R + T +R - T + R +

-4-2

02

4

CD

C2L6

T -R - T -R + T +R - T + R +

-3-2

-10

12

3

FN

BP4

T -R - T -R + T + R - T + R +

-5-4

-3-2

-10

12

MED

4

T -R - T -R + T + R - T + R +

-3-2

-10

12

TB

L1Y

T -R - T -R + T +R - T + R +

-6-4

-20

2

CD

K8

T -R - T -R + T +R - T + R +

-1.5

-1.0

-0.5

0.0

0.5

1.0

CLPTM

1L

Question: What predicts tumor resistance to chemotherapy?Answer: Depletion of differential transcriptome towards few specific genes

Resistance to vinorelbine in lung cancer

Functionally coherent genes associated with vinorelbine resistance. A. Network representation of the group. Magenta: genes associated with resistance in NEA and likely producing a protein complex (ranked 1, 3, 5, 6, 11, and 18) plus one more gene CLPTM1L (beyond the ranking but also significantly associated, was previously reported as related to cisplatin resistance); yellow: non-commonly expressed genes linked with the NEA genes and less represented in the susceptible tumors, hence most contributing to the association (see Results for more explanation).B. Box-plots of NEA scores for the genes colored magenta in A.

Resistance to vinorelbine in lung cancer: pathways

0 2 4 6 8

0.0

0.2

0.4

0.6

0.8

1.0

GO:0001666 response to hypoxia

Years since surgery

Rel

apse

-free

sur

viva

l

0 2 4 6 8

0.0

0.2

0.4

0.6

0.8

1.0

GO:0005154 EGFR binding

Years since surgery

Rel

apse

-free

sur

viva

l

0 2 4 6 8

0.0

0.2

0.4

0.6

0.8

1.0

GO:0005164 TNF binding

Years since surgery

Rel

apse

-free

sur

viva

l

0 2 4 6 8

0.0

0.2

0.4

0.6

0.8

1.0

GO:0070848 response to growth factor stimulus

Years since surgery

Rel

apse

-free

sur

viva

l

Expression:~20000 genes

Clinical data, e.g.: Estrogen receptor status: +/ –Lymph. node status: 0,1,2,3

Relapse : yes/no and time (days)×

Procedure for gene X phenotype signatures

RELAPSE = γ0P + γ1g1 + γ2g2 + γ3g3 + … + γNgN

Overlap between gene signatures of relapse

From Roepman et al., 2007

Usually the overlap between signatures is negligible. Because e.g.:• Different sub-types of patient population, • Different microarray platforms!

However the main reason is:Dimensionality curse!

Biomarker discovery in network context

The idea:Construct multi-gene predictors with

regard to network context

1. Reduce the computational complexity2. Make marker sets biologically sound

Accounting for network context is taking either:a) network neighbors orb) genes at remote network positions

Ready signature in the networkRELAPSE = γ1EIF3S9+ γ2CRHR1 + γ3LYN + … + γNKCNA5

1240 1254 1424 1536 287 664 801 812 1066 1197 1214 172 237 264 400 732 Z-scoreFAM13A 0 0 0 0 0 0 0 0 1 0 1 1 1 1 1 1 -2.61HEPACAM 0 0 0 1 0 0 1 0 1 1 1 1 1 1 1 1 -2.32OR8J1 0 1 0 0 0 0 1 0 1 1 1 1 1 1 1 1 -2.32ARGFX 0 1 0 0 0 0 1 0 0 1 1 1 1 1 1 1 -2.26DZIP1 1 0 0 0 0 0 0 0 1 1 0 1 1 1 0 1 -2.26GABRR2 0 0 0 0 1 1 0 0 1 1 1 0 1 1 1 1 -2.26LIPC 0 0 0 0 0 0 0 1 1 1 1 0 1 1 1 0 -2.26MADD 0 0 0 0 1 0 0 0 1 0 1 1 1 1 0 1 -2.26MMP27 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 -2.26MMP8 0 0 0 1 0 0 1 0 1 1 0 1 1 1 1 1 -2.26PI3 0 0 1 0 0 0 1 0 1 1 1 1 1 1 1 0 -2.26SCGB1D2 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 0 -2.26SCLY 0 1 1 1 1 1 1 1 0 1 0 0 0 0 0 1 2.26SCN11A 0 1 0 0 0 0 0 0 1 1 1 0 0 1 1 1 -2.26VPS13B 0 0 0 1 0 0 0 0 1 1 1 1 1 0 0 1 -2.26APOC4 0 0 1 0 1 0 1 0 1 1 1 1 1 1 1 1 -2.01DNAH3 0 1 1 1 0 0 0 0 1 1 1 1 1 1 1 1 -2.01NID1 0 0 0 0 0 0 0 0 0 1 1 1 1 0 1 0 -2.01OR10P1 1 1 1 1 1 1 1 1 1 0 0 1 0 0 1 0 2.01OR4D11 0 0 0 0 0 0 0 0 0 1 1 0 0 1 1 1 -2.01PAPPA 1 0 0 1 1 0 0 0 1 1 1 1 1 1 1 1 -2.01PGM3 0 0 0 0 0 0 0 0 1 0 0 1 1 1 0 1 -2.01SLC22A4 0 1 1 0 0 0 0 1 1 1 1 1 1 1 1 1 -2.01WDR52 0 0 0 0 0 0 0 0 1 1 1 1 1 0 0 0 -2.01ZNF445 0 0 0 0 0 0 0 0 0 1 0 1 1 1 1 0 -2.01ACTL8 1 1 1 0 1 0 1 1 0 0 0 0 1 0 1 0 1.90ADAMDEC1 0 1 0 0 1 0 0 1 1 0 1 1 1 1 1 1 -1.90BAT2L2 0 0 0 0 1 0 0 0 0 0 1 1 1 1 0 1 -1.90BRCA1 0 0 1 1 0 0 0 0 1 1 1 1 1 0 1 0 -1.90

Chemotoxicity in lung cancer treatment(work with J. Hasmats, H. Green, J. Lundeberg)

With the set of 16 patients, there was not enough statistical power to detect toxicity-associated alleles individually (only 1 gene had a p-value <0.01)

High toxicity Low toxicity

Question: Are there any high-order interactions between toxicity-associated variants?Answer: …

Intra-connectivity of setbest32.top_320.z_1.5

Null.neg Real.neg Null.pos Real.pos Null.rnd Real.rnd Null.top Real.top

-2-1

01

23

4

best32

Null.neg Real.neg Null.pos Real.pos Null.rnd Real.rnd Null.top Real.top

02

46

best16

Neg: variant prevalent in “Low”;Pos: variant prevalent in “High”;Top: Union of Pos & Neg;Rnd: Same size of gene set a s in Top, but test z-score of sign ignored

NULL: Same size of gene set as in Pos, Neg, Top, Rnd, but genes are replaced with random network genes

Enrichment of intra-connectivity of sets of genes with contrast variants over “Low/High”

Mutations accumulated in somatic genomes of cancer cell lines

(Pelin Akan et al., Genome Medicine 2012)

Question: Do mutations within a cancer genome behave like a quasi-pathway?Answer: Yes.

Question: How similar are mutation patterns in different cancers?Answer: A lot, but only for drivers and at the pathway level.

Experimental perturbations of syndecan-1 in cancer cells(T. Szatmari et al., PLoS One, 2012)

Lines:Red: depletionBlue: enrichment

Question: Second-order downstream targets of syndecan modulation?Answer: Segments of cell cycle etc.

Decomposing biological context

rPLC = 0.88

rPLC = 0.95

rPLC = 0.76

Common

Develomental

Dioxin-enabled

ANOVA (Analysis Of VAriance):

Look at F-ratios:

Signal of interest /Residual (“error”) variance

Accounting for edge features:dioxin-enabled vs. dioxin-sensitive links

Andrey Alexeyenko, Deena M Wassenberg, Edward K Lobenhofer, Jerry Yen, Erik LL Sonnhammer, Elwood Linney, Joel N Meyer Transcriptional response to dioxin in the interactome of developing zebrafish. submitted.

FunCoup, inhibitory

Network analysis:how to succeed?

• Analyze prioritized candidates (from genotyping, DE, GWAS…) rather than any genes.

• Do not lean on single “interesting” network links. Employ statistics!i.e.

“concrete questions” => “testable hypotheses” => “concrete answers”

The amount of information in known gene networks is enormous.

Let’s just use it!

Acknowledgements• Simon Merid• Ashwini Jeggari• Darya Goranskaya• Pan Lu

• Erik Sonnhammer– Martin Klammer– Sanjit Roopra– Ted McCormack– Oliver Frings

• Jonathan Prince • Yudi Pawitan

– Setia Pramana– WooJoo Lee

• Joakim Lundeberg• Pelin Akan• Ingemar Ernberg• Per Kraulis

Network enrichment analysis: applications

N links_real = 6N links_expected = 1.00

Standard deviation = 1.25Z = 3.97P-value = 0.0000356

Question:Does gene expression in MAT230414 relate to “response to tumor cell”?

Pathway characterization Detection of driver mutations Coherence of genome alterations

N links_real = 55N links_expected = 37.05

Standard deviation = 3.59Z = 3.59P-value = 0.00016

Question:Could copy number alteration in EHR in HOU501106 lead to changes of its transcriptome and proteome?

N links_real = 0N links_expected = 1.05

Standard deviation = 0.80Z = -1.31P-value = 0.905

Question: Are CNA in HOU501106 coherent?

State-of-the-art method to beat: Observational science

Pathway view on the set of toxicity-associated alleles

The analysis detected more significantly enriched pathways than for the negative control gene sets of the same size (215 vs. 139; p0 < 0.001; FDR<0.05). More specifically, many thus found pathways were associated with cancer, apoptosis, cell division etc.

Red node: list of top 50 genes with most contrast allele patternsGrey node: negative control listYellow: enriched/depleted pathwaysEdge width: no. of gene-gene links in the networkEdge opacity: confidenceGreen edges: enrichmentRed edges: depletion

Edges produced by less than 3 list genes are not shown0

Network enrichment analysis: what to use?

• R package NEA: Alexeyenko A, Lee W, … Pawitan P (2012). Network enrichment analysis: extension of gene-set enrichment analysis to gene networks. BMC Bioinformatics

• Perl software : Simon Merid et al., to be published.• C++, the “crosstalk” tool : Ted McCormack et al., to be

published

• Last but not least: http://FunCoup.sbc.su.se

Decomposing biological context

rPLC = 0.88

rPLC = 0.95

rPLC = 0.76

Common

Develomental

Dioxin-enabled

ANOVA (Analysis Of VAriance):

Look at F-ratios:

Signal of interest /Residual (“error”) variance

Accounting for edge features:dioxin-enabled vs. dioxin-sensitive links

Andrey Alexeyenko, Deena M Wassenberg, Edward K Lobenhofer, Jerry Yen, Erik LL Sonnhammer, Elwood Linney, Joel N Meyer Transcriptional response to dioxin in the interactome of developing zebrafish. submitted.

HyperSet: the network enrichment analysis online

resourcesince 2013 at http://scilifelab.se