joseph powell - using the transcriptome to determine the genetic mechanisms of disease...

Common Diseases Epistasis Variance Heterogeneity

Using the transcriptome to determine the geneticmechanisms of disease susceptibility

Joseph Powell

Centre for Neurogenetics and Systems Genomics

Winter School in Mathematical and Computational Biology

Brisbane - 7th July 2014

Joseph Powell Mechanisms of disease susceptibility


Causes of Death - Common Diseases



Causes of Common Diseases - Genetics + Environment

Environmental factors

Genetics



How Do Mutations Influence Disease Susceptibility?

Most genetic variation underlying disease susceptibility is locatedin regulatory regions



Move To Molecular Phenotypes - Systems Genetics



By what means do most DNA variants alter cellular behaviour andultimately contribute to differences in disease susceptibility?



Genome-wide epistasis Variance heterogeneity



What is epistasis?

Definition

The effect on the phenotype caused by locus A depends on thegenotype at locus B....

Epistasis has been reported in model organisms through artificialgene knockouts, artificial line crosses and hybridization.

Our aim was to systematically search for instances of epistasisamongst common variants for genetic variation that has arisen innatural populations.



Gene expression

Transcription measured forthousands of genes

Typically heritable

Loci (eQTLs) commonlyhave very large effect sizes

Good candidates to searchfor epistasis



Discovery population

846 healthy individuals

528,509 autosomal SNPs

RNA measured forperipheral blood (IlluminaHT-12v4.0)

Expression levels for 7,339probes



Computational analysis

Exhaustive SNP x SNPtesting for each probe

epiGPU software and GPUclusters

8 d.f. F-test

Over quadrillion tests



Design Of The Analysis

SNP x SNP 2 dimension analysis Epistatic genotype by phenotype

Example of genotype by phenotype epistasis map

AA

Aa

aa aa

Aa

AA

0.0

0.2

0.4

0.6

0.8

1.0

SNP1 SNP2

RNA levels



Correction for multiple testing

Apply a family-‐wise error rate of 5%

Permutation and Bonferroniused to give 5% FWER

Permutation: Single probeFWER; T∗ = 2.13x10−12

Correct for 7,339 probes

Experiment wide FWER;T ∗ /7, 339 = 2.91x10−16



Removing artifacts

For SNP pairs that passedthe 2.91x10−16 threshold

All 9 genotypes classespresent

Minimum class size of 5

No LD between SNP pairs(r2 < 0.1 and D ′2 < 0.1)

No single loci additive ordominance effects

11,155 pairs carried forward



Testing for non-additive effects

Nested ANOVA for 11,155pairs

Contrasts full genetic (8d.f.) vs marginal effects (4d.f)

Thus, testing for thecontribution of epistaticvariance



More correction for multiple testing

Epistatic effects significantat p < 0.05/11, 155 =4.48x10−6

501 SNP pairs withsignificant interaction terms



Replication population

Fehrmann

1,240 individuals(Netherlands)

EGCUT

891 individuals (Estonian)


Genotyped with Illuminaarrays



Replication tests

Replication of 501 epistatichits using two independentcohorts

Same filtering criteria

434 pairs which could betested in both cohorts



Functional work

Analyses to elucidate possiblefunctional mechanisms

Non-synonymous mutations

GWAS and known eQTLoverlap

Genome segmentation

Chromosome interactions

Tissue specific transcriptionregions



Q-Q plots of interaction p-values from replication

The top panel shows all 434discovery SNPs that were testedfor interactions.

The bottom panel shows thesame data as the top panel butexcluding the 30 most significantinteractions

Matched direction of epistaticeffects p = 5.56x10−31

●

●

●●●

●●●

●●

●●

●

●●

●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●

●

●

●

●

●

●●●

●

●●

●●●

●●●●

●●●●●●●●●

●●●●●

●●●●●●●

●●●●●●●●●●●●

●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

0

50

100

150

0

1

2

3

4

All

Below

Bonferroni threshold

0 1 2 3Expected

Obs

erve

d



Interactive Figure: http://kn3in.github.io/detecting_epi/

Yellow dots: A Transcript Grey dots: Epistatic SNPs

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

● ●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●


http://kn3in.github.io/detecting_epi/


Example: Epistatic Control Of CAST Regulation

CAST: Associated with numerousrheumatic diseases

It’s regulation controlled by manyepistatic SNPs throughout thegenome

Novel genetic control of regulation

potentially underlying disease

susceptibility

Chromosomes

Position of CAST Gene Epistatic

interactions



Functional Mechanisms - Chromosomal Interactions

Chromosome interactionsidentified in K562 cell linesusing Hi-C and ChIP-seq

Red lines = N SNPs within20kb, 500kb, 2Mb and 10Mbof known interacting regions

Histograms = null distributions

0

250

500

750

0

250

500

750

0

250

500

750

0

250

500

750

5Mb

1Mb

250kb10kb

0 10 20 30 40Number of chromosome interactions

Fre

quen

cy



Conclusions

The first evidence for multiple instances of segregating commonpolymorphisms interacting to influence human traits

Novel computational and statistical frameworks can identify andreplicate epistasis

New knowledge of the control of transcription of many genesimplicated in common disease



Variance Heterogeneity

Searching for veQTL

There is evidence across several species for genetic control ofphenotypic variation of complex traits

This means that the variance of a phenotype is genotype dependent

Understanding genetic control of variability is important inevolutionary biology, and medicine



Variance eQTL

What causes them?

Epistatic interactions

Gene by Environment Interactions

Other ..... ?



846 healthy individuals

6,011,184 Imputed(autosomal) SNPs


Expression levels for 17,994probes



Double Generalized Linear Model

y = µ+ xb + e

e ∼ N(0, σ2e )

Model to detect variance and

mean effects

e ∼ N(0, σ2ei ); log(σ2

e ) = c + xv

y ∼ N(µ+ xb, exp(c + vx))



For statistical analysis, theinformational content is the same ifdata is transformed by a monotonicfunction

Provided the three means aredifferent, there is always amonotonic transformation that willtend to equalize the three variances

However, it’s relevance depends on

the biological scale



Original and normalised scales



Scale

Initial scale = Z-score frominverse-normal transformation

Adjust original scale using transformationfunctions provided in Sun et al. (2013)

σ2(x) = f (m) = am2 + bm + c (1)

Express variance of x as a function of its meanmx , f (mx )

y = g(x)∞∫

dx√f (x)

(2)

Re-test using dglm

The aim is to remove main driven variance effects

Pictorial representation of four different non-linear

changes to the scale. They represent four classes of

monotonic transformation (Sun et al AJHG 2013).



Discovery results

Cis - 2.55x10−7

747 veQTL (Potentially havesome main effects)

Only 88 have Pmain > 0.01

Transformations remove theeffects of 655/659 (with maineffects)

Transformations remove theeffects of 1/87 (without maineffects)

Trans - 3.39x10−12

1412 veQTL (Potentially havesome main effects)

619 have Pmain > 0.01

Transformations remove theeffects of 732/793 (with maineffects)

Transformations remove theeffects of 21/619 (withoutmain effects)



Fehrmann

1,240 individuals(Netherlands)

EGCUT

891 individuals (Estonian)


Genotyped with Illuminaarrays



Replication of 7,768 veQTLhits using two independentcohorts

DGLM

Also Bartlett’s and Levene’stests



Replication Meta-analysis

Just interested in the variance only veQTL

At FDR = 0.05

89 % of cis-veQTL

78 % of trans-veQTL

Same effect direction;

85 % cis and 77 % trans



Ripple effect across a network

●●

●

●

●●

●

●

−5.0

−2.5

0.0

2.5

0 1 2factor(rs3821224)

ILM

N_1

7983

08

Predicted TF binding sites tagged by expression probes. Blue lines havevariance heterogeneity with rs3821224 (p < 0.05)

Expression levels for a probe in AHSA2 by genotype class of rs3821224



GWAS - veSNP overlap

Disease / Trait N veSNPs in GWAS catalogueAsthma 1Autism spectrum disorder 2Bipolar disorder 2Cholesterol, total 3Endometriosis 1HDL cholesterol 2Height 4Inflammatory bowel disease 2LDL cholesterol 6N-glycan levels 4Obesity-related traits 5Parkinson’s disease 6Rheumatoid arthritis 3Schizophrenia 8Sudden cardiac arrest 2Systemic lupus erythematosus 3Triglycerides 7Type 1 diabetes 3Urate levels 2



Conclusions

Computational and statistical frameworks can identify andreplicate variance heterogeneity for RNA transcripts.

Variance (only) eQTL represent another form of genetic control forphenotypes.

Functional characterization of variance loci suggests a number ofpossible mechanisms that underlie variance heterogeneity.



Acknowledgements

Centre for Neurogenetics and Systems Genomics

Gib Hemani

Kostya Shakhbazov

Jian Yang

Allan Mcrae

Peter Visscher

University of Groningen

Harm-Jan Westra

Lude Franke

Dalama University

Lars Ronnegard

QIMR Berghofer

Nick Martin

Anjali Henders

Grant Montgomery

Georgia Institute of Technology

Greg Gibson

University of Tartu

Andres Metspalu

Tonu Esko

Visit us at www.complextraitgenomics.com


joseph powell - using the transcriptome to determine the genetic mechanisms of disease...

Science

instances of epistasis

common variants

analysis snp x snp

snp pairs r2

multiple testing epistatic

pairs contrasts

genetic variation

fwer permutation