1. New Statistical Methods for Family-Based Sequencing Studies
Correcting for confounding introduced by batch effects and imputation in family-based designs with whole genome sequence data
Ellen M. Wijsman
New Statistical Methods for Family-Based Sequencing Studies
BIRS August 5-10, 2018

2. Motivation
Interest in rare variation motivates family designs and next generation sequencing (NGS)
Focus here on association analyses (LMM) accounting for kinship
NGS, especially whole genome sequencing (WGS) is expensive
Missing data is the norm: not all family members can or will be sequenced
Subjects sequenced may favor cases/extreme phenotypes
May want to re-use of existing data from other studies, e.g., as controls
Genotype imputation can augment observed data
WGS takes time & may be carried out at different locations
Additional samples may be added later

3. Challenges
Multiple potential factors can lead to biased association tests/confounding
Sequencing Platform (HiSeq 2000, HiSeq X)
Sequencing Center (BROAD, Baylor, Wash U)
Software & protocols (Atlas, GATK, filters)
Reference materials (Genome ver., SV reference files)
Batches within Center
Versions of software, reference info, etc.
Differential genotype quality
Read depth variation
Genotype quality choices differ in calling
Observed vs. imputed genotypes
Raw reads
(fastq files)
Aligned reads
(BAM files)
Called genotypes (VCF files)
Data analysis

4. The Alzheimer’s Disease Sequencing Project (ADSP) – family data (WGS)
Investigate the role of rare variation in Late Onset AD (LOAD)
Learn about importance of coding vs. non-coding variation in LOAD
Big enough to learn something about new proposed analysis approaches
110 multiplex families, selected to avoid families with APOE e4/e4 subjects
~3400 total subjects, ~30% with some genotype data
2-5 generations/family
40% White, 60% Dominican Republic Hispanic
Supported by the NIH National Institute on Aging and the National Human Genome Research Institute
Data available to qualified Investigators via dbGaP/NIAGADs

5. ADSP family data Phases Subjects/samples/trait data Genomic data
Discovery, GRCh37 ( )
513 cases : 65 controls
Extension, GRCh38 ( )
30 cases : 221 controls
Subjects/samples/trait data
ADGC consortium (19 sites)
CHARGE consortium (6 sites)
NCRAD
Genomic data
SNP arrays: individual sites
WGS: national sequencing centers
Centers
Baylor University
Broad Institute
Washington University
Platforms
HiSeq 2000 (Discovery)
HiSeq X 10 (Extension)
Genotype calling

6. Genotyping in ADSP families
Point out subjects with phenotype data that might get good imputed data
27 Subjects
12 GWAS No genos
6 WGS+ 6 no WGS 7 no phenos
GWAS SNPs
chromosome
WGS SNVs

7. Imputation from Inheritance Vectors (IVs)
MORGAN/gl_auto & ~6000 SNPs
uses MCMC or exact computation: pedigrees can be very large, and (somewhat) complex
obtains samples of IVs at the SNPs in intact pedigrees
IV1: [ ]
IV2: [ ]
IV3: [ ]
Cheung et al 2013
AJHG 92:

8. Imputation from Inheritance Vectors (IVs)
Use sampled IVs for imputation from WGS with GIGI, averaging across multiple sampled sets of IVs
Sample an IV at the position of a WGS variant, given IVs at flanking SNPs
Realize genotypes at the WGS variant, given observed data
Average across sampled realizations
Result for subject j:
Association testing with LMM (GEMINI), average imputed genotypes across sampled IVs (dose)
infer
WGS geno a A a a A A
Cheung et al 2013
AJHG 92:

9. Initial association analysis: discovery data with imputed genotypes
Whites - genomewide
Oh my gosh – what is going on?
When your QQ plot looks like a giraffe, there is clearly a problem somewhere!
AD ~ SNV + PCs

10. It took awhile to figure out...
The problem was: There wasn’t just one problem.

11. Simplify the data Use reduced number of WGS genotypes
Reduce computational burden
Focus on variants that have “behaved” in the past
Avoid ultra-rare variants
Use previously vetted variants
Use ~250,000 variants from the WGS also on SNP arrays

12. Association testing with imputed data: discovery data
WGS+ GWAS
Initial QQ plot: Hispanic data
No Obs. Data
GWAS only
WGS+GWAS: there is *some* missing data since not every WGS genotype is called
AD ~ SNV + PCs
0.0
Imputation Deviance
0.6

13. Measuring imputation (un)certainty
Original purpose: provide a “vote” for best genotype when merging imputation from two sources
High variance among genotype probabilities for a SNP indicates high certainty: V(p) =
observed: V(p) = 1/3
no info: V(p) = 0
incomplete info:
Saad & Wijsman, 2014 GEPI

14. Imputation deviance
Imputation certainty is a data-driven proxy for accuracy
Transform to imputation deviance: average over markers for a subject:
Range: 0 (best) to 1 (worst)
Low deviance implies high accuracy 1
Imputation deviance

15. Association testing: discovery sample
Original QQ plot
+Imputation deviance
Fixing the bad QQ plots
Cause: Confounding
~8:1 case:control ratio with WGS
Higher fraction of controls with (imperfect) imputed WGS
Solution: adjust for
admixture (or PCs)
imputation status
imputation deviance A B
+Imp. deviance & status
All Hispanic families
cases vs. controls C
AD ~ SNV + PCs A bad
+ imputation deviance B better
+ imputation status C best !

16. Association: Discovery+extension
Whites: WGS subjects only
Extension data:
New sequencing machines
Called on new sequence build
Heavily tilted towards controls
Discovery+extension re-called together
Sequencing platform: a new source of error needs additional correction
Cov: PCs

17. Association: Discovery+extension
All whites (obs. & imputed)
All whites (obs. & imputed)
Cov: PCs, imp.dev, imp.status
Cov: PCs, imp.dev, imp.status, platform
Not perfect yet...

18. Summary New data, new problems
Multiple sources of differences in data sources can lead to confounding in analysis of NGS in pedigree data
Simple batch effects are not always sufficient corrections
A continuous measure related to the final product is helpful
This is issue is going to become more common
We may have identified an approach to deal with data that is pertinent in other situations
case-control samples with population imputation – varies across the genome.

19. Thank you!