An introduction to Genetical Genomics and Systems Genetics to better understand cancer and other com

1
An introduction to Genetical Genomics and Systems
Genetics to better understand cancer and other
complex diseases

• Ina Hoeschele
• September 2006

2
Genetic base of complex disease traits

• Understanding the genetic determination of
  complex, disease-related traits is long-standing
  goal
• Some standard approaches
• Mapping of quantitative trait loci (QTL) via
  linkage or association mapping
• In human populations
• In animal models of specific diseases
• Collaboration with Professor Miller (Cancer
  Biology) to identify candidate genes responsible
  for the observed phenotypic variance in lung
  tumor incidence of adult mice following in utero
  exposure to chemical carcinogens
• Still no ready to use software for some animal
  models (AIL, RIX)

3
Genetic base of complex disease traits

• Some standard approaches
• Gene expression studies
• Global gene expression profiling of individuals
  with a disease and individuals without the
  disease, or two groups of individuals with
  different sub-types of a disease (observational
  data)
• List of differentially expressed genes (candidate
  genes for QTL)
• Identification of differentially expressed and/or
  differentially regulated gene networks
• Collaboration with Professor Miller
  Identification of time dependent alterations in
  signaling networks that drive tumor progression
  from benign to advanced

4
Genetic base of complex disease traits

• Systems Biology Reconstruction of cellular
  networks involving genes, proteins, metabolites
• Gene networks
• phenomenological, not physical
• network of interactions or regulations among
  genes
• provide valuable information about the genetic
  architecture of complex diseases
• classical genetics concepts such as dominance and
  epistasis can be understood in terms of gene
  networks and their properties
• important practical applications, identification
  of drug targets

5
(No Transcript)
6
Gene networks from observational data

• This network can be obtained using linear
  graphical methods
• This is an interaction network edges between any
  pair of genes that directly interact with each
  other
• the edges are undirected - it is not a causal or
  regulatory network
• it does not tell us which gene regulates which
  other gene(s)
• We compute such a network as an Undirected
  Dependency Graph

7
Gene networks from observational data

• Undirected Dependency Graph UDG
• Based on partial correlations
• Corr(gene A, gene B) 0.7, Corr(A,C) 0.4
• A ? B ? C Corr(A,C B) 0
• Sometimes, we can find regulation
• A ? C ? B A and B jointly regulate C
• A and B do not regulate each other
• Corr(A,B C) ? 0 BUT Corr(A,B) 0
• Observational data
• Power and Sample Size
• Several simulation studies 100 genes, 200
  edges, samples sizes 50 200 power lt 0.20 to
  0.60

8
Causal, regulatory gene networks

• Also represented by graph with gene nodes but now
  edges are directed
• A ? B gene A regulates gene B
• Two approaches
• Time series experiment
• Specific perturbation experiment
• One-at-a-time specific perturbations in same
  genetic background (several genes are knocked
  down, one at a time, RNA interference)
• Multifactorial perturbations genetical
  genomics or expression genetics (Jansen and
  Nap 2003 Trends in Genetics)

9
Design and Data

• Specific Perturbation Experiment

10
Design and Data

• Multi-factorial perturbation experiment
  Genetical Genomics (M marker tested genome
  location)

11
Genetical Genomics / Systems Genetics

• Segregating population of n x 100 individuals
  (n1,2,3,), each individual is
• DNA marker genotyped (genome-wide)
• Expression profiled (genome-wide)
• Phenotyped for disease-complex related traits

We need to find out which DNA markers are
expression QTL (eQTL)
12
Genetical Genomics / Systems Genetics

• Goal causal, regulatory gene network
• Identification of DNA markers affecting the
  expression profiles (etraits) of genes eQTLs
• Identification of regulator-target gene pairs
  from
• set of genes physically located in an eQTL region
  (candidate regulators)
• set of genes affected by the eQTL region
• Construction of Encompassing Network (EN)
• Search for set of optimal networks within EN

13
Genetical Genomics / Systems GeneticsIdentificat
ion of expression QTLs

• Identification of DNA marker influencing an
  etrait via linear regression
• Eig ?g bgXik eig Xik 0 (AA) / 1
  (BB)
• H0 bg 0
• Standard approach genome-wide search of each
  etrait separately
• Mapping of principal components based on PCA of
  all genes or subsets of genes obtained by cluster
  analysis

14
Genetical Genomics / Systems GeneticsGenetic
Mapping of Expression QTLs

• Cis- and trans-eQTL mapping
• Cis-eQTL
• eQTL affects the expression of a gene located at
  the eQTL eQTL is a DNA polymorphism in the
  promoter of the gene
• Test only the marker(s) closest to the gene (not
  genome-wide)
• Cis-Trans-Regulation
• Test the effects of any cis-eQTL on other genes
  cis-eQTL ? Gene A ? Gene B
• cis-eQTL will affect expression of gene B

15
Genetical Genomics / Systems GeneticsGenetic
Mapping of Expression QTLs

• Cis- and trans-eQTL mapping
• Trans-eQTL A ? B ? trans-eQTLA
• Coding region polymorphism in gene A which
  regulates gene B
• Test jointly the effect of candidate regulator
  gene (kA) and its nearest DNA marker on
  expression of target gene (gB)
• Eig ?g b1gXik b2gEik ( b3gEik?Xik) eig
  
• Xik 0/1
• Intersection-Union test to identify cases where
  b1g and b2g are both non-zero

16
eQTL overlap for SPA, PC-mapping, Cis-mapping and
Trans-mapping
7
13
21
70
83
2
1
3
SPA
Cis-mapping
24
3
8
3
8
87
16
49
2
PC-mapping
Trans-mapping
17
Genetical Genomics / Systems GeneticsIdentificat
ion of regulator-target pairs
eQTL
Regulators genes physically located in eQTL
Targets genes whose etraits are affected by
eQTL
18
Genetical Genomics / Systems GeneticsIdentificat
ion of regulator-target pairs

• Cis- versus cis-trans regulation

19
Genetical Genomics / Systems GeneticsIdentificat
ion of regulator-target pairs

• Cis-trans versus trans regulation

20
Genetical Genomics / Systems GeneticsIdentificat
ion of regulator-target pairs

• Trans regulation (is gene A a trans-regulator)
• Intersection-Union test for b1D and b2D

21
Regulator-target pairs - SPA
23
15
15
6
41
22
Regulator-target pairs - Cis mapping
14
24
41
6
15
23
Regulator-target pairs PC mapping
62
8
9
16
5
24
Regulator-target pairs Trans-mapping
25
Genetical Genomics / Systems GeneticsEncompassin
g (Directed) Network EDN

• EDN is obtained by combining all retained
  regulator-target pairs
• cis-eQTL ? target gene
• cis-regulated gene ? target gene (cis-trans
  regulation)
• trans-eQTL ? target gene
  regulator gene ? target gene (trans-regulation)
• Next step constrained network search within the
  EDN

26
Yeast Data Encompassing Network

• Yeast data of Brem and Kruglyak (2005)
• 112 haploid offspring of cross between wild and
  laboratory strain of yeast
• Expression-profiled for 6000 genes
• DNA marker genotyped for 3000 markers
• Used 4589 etraits and 2956 markers
• Encompassing network
• 28,609 regulator-target pairs
• 4,274 gene nodes
• 2118 gene regulators, 4116 gene targets

27
Yeast Data Encompassing Network

• Encompassing network
• Regulator with most targets (PHM7) 468
• Target with most regulators (YLR152C) 32
• Confirmed regulators
• Amn1 top cistrans regulator with 408 targets
• MAK5 110 trans targets
• GPA1 60 targets (half trans)
• Heme dependent transcription factor HAP1 141
  targets (100 cistrans)

28
Genetical Genomics / Systems GeneticsGene
network reconstruction

• Popular tool Bayesian Network (BN)
• Represents conditional independence A?B?C
• Suitable for noisy data
• Search among equivalence classes (equivalent
  models cannot be distinguished based on available
  data) A?B A?B or A?B
• Limited to DAG directed, acyclic graph no
  cycles or feedback loops
• Time dimension A1?B2?C3?A4? dynamic BN
• Usually, the expression data are discretized

A
B
C
29
Genetical Genomics / Systems GeneticsGene
network reconstruction

• Our tool Structural Equation Modeling (SEM)
• Represents conditional independence A?B?C
• Uses continuous expression data with normality
  assumption robustness?
• Suitable for noisy data
• Can model DCG directed cyclic graph
• Search among models, not equivalence classes
• Edge directions are fixed
• Among two equivalent models with different
  numbers of edges (possible in DCG), we prefer the
  sparser
• No efficient algorithm for identification of
  equivalence classes

30
Genetical Genomics / Systems GeneticsGene
network reconstruction

• Structural Equation Modeling (SEM)

yn expression data, xn eQTL genotype codes
Regulator gene A
bBA
fC2
bBC
Target gene B
Regulator gene C
hBA1
eQTL 2
fB1
eQTL 1
kB12
31
RESULTS

• SEM widely used in econometrics, sociology and
  psychology (confirmatory, not exploratory)
• Typically applied to at most 10s of variables
• Available in many software packages (Lisrel, Mx,
  ) but not feasible for genome-size data
• Own implementation based on C can handle 100s
  of genes (in Genetical Genomics context)
• SEM applied to sub-network of Yeast Encompassing
  Network
• 265 genes, 241 eQTL, 832 gene-gene edges, 640
  eQTL-gene edges, cycle with 168 genes
• Sparsified network with 475 gene-gene edges and
  468 eQTL-gene edges

32
Yeast network
EDN 265 genes, 241 QTLs, 832 gene ? gene edges,
and 640 QTL ? gene edges After sparsification
475 gene ? gene edges and 468 QTL ? gene edges
33
RESULTS

• SEM applied to artificial gene expression data
  (known network structure)
• 10 data sets with different, random network
  topologies, 100 genes, 100 eQTLs
• mRNA levels simulated with non-linear ODE
  (Gepasi)
• On average 148 gene-gene and 123 eQTL-gene edges
• EN with 360 gene-gene and 301 eQTL-gene edges
• On average 42 genes in cycles (1-3)
• False Discovery Rate ( wrongly identified edges
  / total identified edges) 0 - 15.
• Power ( edges correctly inferred / total edges
  in the true network) 72 - 100

34
Systems Genetics

• The merger of Systems Biology with the study of
  genetic variation
• The integration and anchoring of
  multi-dimensional data-types to underlying
  genetic variation (transcriptomic, phenomic,
  metabolomic )

35
Systems Genetics

• Using a segregating population, we can
• Reconstruct a network of direct relationships
  among various phenotypic measurements related to
  a disease complex
• Reconstruct a causal gene regulatory network
  (with DNA marker and expression profiling)
• Combine the above into a causal network of genes
  and disease phenotypes
• Determine the extent of genetic control of
  metabolomic variation (with additional
  metabolomic profiling) and transcriptional
  control of metabolic reactions

36
Systems Genetics

• Using a segregating population, we can
• Reconstruct a network of direct relationships
  among various phenotypic measurements related to
  a disease complex
• e.g., Cardiovascular System
• Bone Fragility (morphology, mineral content,
  mechanical properties, body composition)

37
Systems Genetics

• Our prospects for investigating the complex
  interactions between gene variants, disease, and
  the environment will be significantly improved
• Our understanding of the gene and metabolic
  regulatory circuitry and its relationship with
  disease phenotypes will be greatly enhanced

38
Questions / Comments
39
Gene networks from observational data

• Undirected Dependency Graph UDG
• Observational data tumor and normal samples
  with measured gene expression for all genes
• Method A
• Start with a network that has an edge between any
  pair of genes which are significantly correlated
  (many of these interactions are indirect through
  other genes)
• For each pair of genes, compute Corr(G1,G2 Gk),
  where k denotes any gene other than G1 and G2,
  retain only those edges with significant 1st
  order partial correlation
• For each pair of genes, compute Corr(G1,G2 Gk,
  Gm), where k and m are any two genes other than
  G1 and G2, retain only those edges with
  significant 2nd order partial correlation
• Etc.

40
Gene networks from observational data

• Undirected Dependency Graph UDG
• Method B
• Estimate the covariance matrix of all genes and
  invert this matrix.
• genes gtgt observations, sample covariance
  matrix does not have inverse
• use of special shrinkage estimator
• Inverse ? matrix of partial correlations
  Corr(G1,G2 all other genes)
• Begin with a network having edges between any
  genes with significant partial correlation
• For each pair of genes, compute Corr(G1,G2),
  retain only those edges with significant simple
  correlation
• For each pair of genes, compute Corr(G1,G2 Gk),
  retain only those edges with significant 1st
  order partial correlation
• For each pair of genes, compute Corr(G1,G2 Gk,
  Gm), retain only those edges with significant 2nd
  order partial correlation
• Etc.