Transcript Evaluating 1x coverage NGS for GWAS in 12,000 individuals

Genotype Phasing and Imputation in 1x Sequencing
Data
Warren W. Kretzschmar
DPhil Genomic Medicine and Statistics
Wellcome Trust Centre for Human Genetics, Oxford, UK
Supervisor: Jonathan Marchini
Major Depression
• Commonest psychiatric
disorder and the
second ranking cause
of morbidity worldwide.
• Affects 1 in 10 people
in their lifetime.
• Estimates of
heritability range
between 30-40%.
Top Ten causes of DALYs
DALY : Disability adjusted life year
: number of years lost due to ill-health, disability or early death
Lower
respiratory
infections
Other
unintentional
injuries
Cerebrovascular
disease
Chronic
obstructive
pulmonary
disease
Major
depressive
disorders
Violence
Ischaemic
heart
disease
Diabetes
mellitus
Road traffic
accidents
Alcohol use
disorders
Genetics of Major Depression
Major Depressive Disorder Working Group of the Psychiatric GWAS
Consortium (2012). A mega-analysis of genome-wide association studies
for major depressive disorder. Molecular Psychiatry 18.4:497-511.
Study Design
• Unrelated Europeans
• 9240 cases
• 9519 controls
• 1.2 million SNPs
Hypotheses
• Depression has
heterogeneous
environmental and
genetic causes
• Depression is a complex
trait with genetic
components of small
effect size
CONVERGE
(China, Oxford and VCU Experimental Research
on Genetic Epidemiology)
59 hospitals, 45 cities, 21 provinces.
Genetically Homogeneous : All subjects are
female and their grandparents are Han Chinese
6,000 cases : typically severe affected: 85%
qualify for a diagnosis of melancholia by DSMIV. >25% reported a family history of MD in one
or more first-degree relatives
6,000 controls : patients undergoing minor
surgical procedures.
Extensive Phenotyping : primary disorder of major depression, common comorbid
disorders (e.g. generalized anxiety disorder, panic disorder), within disorder symptoms (e.g.
suicidal ideation), disorder subtypes (e.g. melancholia, dysthymia), possible
endophenotypes (e.g. neuroticism) and a range of risk factors (e.g. child abuse, stressful life
events, social and marital relationships, parenting, post-natal depression, demographics).
Sequencing : mean depth 1.7X using lllumina HiSeq at Beijing Genomics Institute
Current status Sequencing finished. We have data on 12,000 samples. For now we have
only considered ~13M sites polymorphic 1000 Genomes Asian samples. Analysis ongoing…
Sequence analysis pipeline
Phase 1: genotype likelihood estimation
One sample at a time
Genotype
likelihoods
Raw reads
Mapping
Stampy
Phase 2: phasing and imputation
All samples together
My focus!
48 TB
350 GB
Duplicate
marking
Picard
Base quality GATK
recalibration
Genotype
likelihood
estimation
Phasing and 2.7 CPU
imputation
years
5 CPU
years
650 GB
SNPTools 4.6 CPU
years
Genotype
likelihoods
Genotype
probabilities
GENOTYPE PHASING AND
IMPUTATION
Genotype Phasing
Example SNP chip data
Unphased:
G/G A/T A/A T/T G/T A/T T/T A/A G/G G/C
After Phasing
Hap 1:
G
A
A
T
T
T
T
A
G
C
Hap 2:
G
T
A
T
G
A
T
A
G
G
Phase-informative Sites
Genotype Imputation from Haplotypes
J Marchini and B Howie. Nature Rev. Genet. 2010
GENOTYPE LIKELIHOODS
What is a Genotype Likelihood?
Genotype likelihoods (aka GL) are defined on a site by site basis.
GLs are conditional probabilities.
Genotype Likelihood = Pr( R | G )
R = Reads; also known as the “observed data”
G = Genotype; usually one of ref/ref, ref/alt, alt/alt
How are Genotype Likelihoods Useful?
Genotype likelihoods allow us to quantify how much the reads support
each possible genotype independent of other information.
To determine the most likely genotype call, we need a genotype
probability.
Genotype Probability = Pr ( G | R ) proportional to Pr( R | G ) * Pr( G )
Pr( G ) = prior probability of G.
May be determined through haplotype phasing and imputation approaches.
Genotype Likelihood Creation with SNPTools
observed
reads
Pr(R|G = ref/ref)
= 0.06
Pr(R|G = alt/alt)
= 10e-6
Pr(R|G = ref/alt)
= 10e-3
Y Wang, J Lu, J Yu, RA Gibbs, FL Yu. Genome Research. 2013
Genotype Phasing using Genotype Likelihoods
Reference Haplotypes
Hap 1:
G
A
A
T
T
A
C
A
G
G
Hap 2:
G
T
A
T
T
A
T
A
G
G
Hap 3:
G
T
A
T
G
A
C
A
G
G
Hap 4:
G
T
A
T
G
A
T
A
G
C
Example GL data
Pr(ref/ref):
Pr(ref/alt):
Pr(alt/alt):
Plausible Haplotypes after Phasing
Hap 5:
G
A
A
T
T
A
T
A
G
C
Hap 6:
G
T
A
T
T
A
T
A
G
G
General MCMC Scheme for Phasing from GLs
When using GLs, haplotype estimation is currently done in an
iterative Markov Chain Monte Carlo (MCMC) scheme
1. Initalize haplotypes for each sample randomly
2. for a predetermined number of iterations
1. for each sample
1. Find a plausible haplotype pair using its GLs and all
other haplotypes as a reference panel
2. Update that sample’s haplotypes with the plausible
haplotype pair
3. Return each sample’s current pair of haplotypes
The Tools/Languages I use
Coding
Emacs
Scripting
Perl with DistributedMake for
pipelines
Statistical Methods
C++
Figure Generation
R
Statistical Analysis & Report
Writing
LaTeX with SWeave
Presentations
PowerPoint or LaTeX
A Bioinformatician’s Best Practices
according to Nick Loman & Mick Watson. Nature Biotechnology. 2013
see also: W. S. Noble. PLoS Computational Biology. 2009
- Understand your goals and choose appropriate methods
- Be suspicious and trust nobody
- Set traps for your own scripts and other people’s
- Be a detective
- You're a scientist, not a programmer
- Use version control software
- Pipelineitis is a nasty disease
- An Obama frame of mind
- Someone has already done this. Find them!
Good Directory Structure
according to W. S. Noble. PLoS Computational Biology. 2009
Thank you. Questions?