Epistasis

02-715 Advanced Topics in Computa8onal

Genomics
 Epistasis

Deni8on: The eect of one locus depends on the genotype

of another locus
Epista8c eects vs. marginal eects
Epistasis for Mendelian Traits

Carlborg & Haley, Nature Reviews Gene8cs 2004

When There is No Epistasis

Two addi8ve
(non-epista8c)
loci

The three lines

run in parallel
Epistasis Example

Dominant epistasis

One locus in a
dominant way
suppresses the
allelic eects of a
second locus
Epistasis Example

Co-adap8ve epistasis

Genotypes that are

homozygous for alleles
of the two loci that
originate from the same
line (JJ with JJ, or LL
with LL) show enhanced
performance.

Almost no marginal
eects: average eect
of JJ, JL, LL do not dier
Epistasis Example

Dominance-by-
dominance epistasis

Double heterozygote
(LS, LS) deviates from
the phenotype that is
expected from the
phenotypes of the
other heterozygotes.

Double heterozygotes
have a lower phenotype
than expected.
 Epistasis

Epista8c eects of SNPs can oWen be detected only if the

interac8ng SNPs are considered jointly
The number of candidate SNP interac8ons is very large
For J SNPs, JxJ SNP pairs need to be considered for epistasis
In general for J SNPs and K-way interac8ons, there are O(JK)
candidate interac8ons
Computa8onally expensive to consider all possible groups of
interac8ng SNPs
For a reliable detec8on of K-way interac8ons, a large sample size is
required
Mul8ple tes8ng problem
 BEAM Overview (Zhang and Liu, 2007)

Bayesian epistasis associa8on mapping for case/control

studies

Bayesian par88oning of the gene8c markers to groups of

markers with/without associa8ons

Use MCMC to learn the par88oning and obtain the posterior

probability of par88ons of markers

It can handle up to ~100,000 markers

 BEAM

Assume the markers in the case group belong to Group 0, 1,

or 2.
Group 0: Markers with no eects D1
Group 1: Markers with marginal eects D2
Group 2: Markers with epista8c eects D3

Markers in the control group U belong to Group 0.

Goal: Learn the par88on of markers I into Groups 0, 1, 2,

given the genotype data.
 Bayesian Marker Partition Model

Model for markers in case group with marginal eects

Assume a Dirichlet() for 1

Marginal likelihood aWer integra8ng out parameters 1

 Bayesian Marker Partition Model

Model for markers in case group with epista8c eects

Markers with no eects in case group, and markers in the

control group
 Bayesian Marker Partition Model

The posterior distribu8on of marker assignment to Groups 0,

1, 2 is given as

The prior is given as

 BEAM: MCMC Sampling

Ini8alize I according to P(I)

Metropolis-Has8ngs (MH) algorithm
Propose to change the markers group membership
Propose to randomly exchange two markers between Groups 0, 1,
2.
 Evaluating Markers for Associations

Use the posterior probability P(I|D,U)

Use B-sta8s8c

Bayes factor under null model

Bayes factor under the alterna8ve model
 Simulation Study

Model 1: two disease

loci with independent
eects
Model 2: disease risk
only when both loci have
at least one disease
allele
Model 3: addi8onal
disease alleles at each
locus do not further
increase the disease risk
 Simulation Study

Other scenarios to be considered

Model 4: three disease loci.
Model 5: mul8ple causal epistasis by a mixture of two two-way
interac8ons. Disease risk if at least one epista8c interac8on is present.
Model 6: six-way interac8on
 Results for Models 1-3
2000 cases, 2000 controls
power

1000 cases, 1000 controls

 Results on AMD Dataset

AMD dataset, 116,204 SNPs genotyped for 96 cases and 50

controls. Posterior probabili8es for having an associa8on
 Sensitivity to Prior Distributions

p1=p2 p1=p2
=0.0033 =0.001

p1=p2 p1=p2
=0.0001 =0.00001

P1: prior probabili8es for marginal eects

P2: prior probabili8es for epista8c eects
 MCMC Trace and Autocorrelation Plot

Simulated
data

AMD data
 Screen and Clean (Wu et al., 2010)

Mul8variate analysis of all SNPs and SNP interac8ons using

lasso

Assume that the SNPs with epista8c eects are likely to have
at least small marginal eect
Select the SNPs with marginal eects.
Consider only those SNPs with marginal eects for epista8c
interac8on.
SNPs will be found to have epista8c eects but no marginal eects
 Screen: Two-stage Lasso for Detecting
Epistasis
Two-stage lasso
Step 1: Apply lasso with no considera8on of epistasis to detect SNPs
with signicant individual eects

Step 2: Apply lasso with pairs of only those SNPs selected in Step 1
Clean: Assessing Significance of Results

Split the data into Stage 1 and 2 datasets

Using Stage 1 data, apply lasso to select SNPs with marginal and
epista8c eects
Using Stage 2 data, assess the signicance of SNPs
Apply the least squared error method and obtain the tradi8onal t-
sta8s8c

Mul8-split method
Randomly split the data into Stage 1 and Stage 2 datasets mul8ple
8mes
Perform the two-way split method for each split
Combine the p-values from each split.
 Advantages and Disadvantages

Reduces the computa8onal burden.

SNPs with epista8c eects oWen do not have detectable

individual (marginal) eects, and many of these associa8on
signals will be missed.

Need to split the data into two parts for Stage 1 and Stage 2
analysis
Screen and Clean
Simulation Results

Solid line: mul8-split,

dojed line: single-
split
 Results

WTCCC data
1,963 cases with T1D and 2,938 controls
Mul8-split screen and clean with 56 random splits