Introduction to Variant Calling

Michael Lawrence

June 24, 2014

Outline

Introduction

Calling variants vs. reference

Downstream of variant calling

VariantTools package

Visualization

Outline

Introduction

Calling variants vs. reference

Downstream of variant calling

VariantTools package

Visualization

Variant calls

DefinitionI A variant call is a conclusion that there is a nucleotide

difference vs. some reference at a given position in anindividual genome or transcriptome,

I Usually accompanied by an estimate of variant frequency andsome measure of confidence.

Use cases

DNA-seq: variants

I Genetic associations with diseaseI Mutations in cancerI Characterizing heterogeneous cell populations

RNA-seq: allele-specific expression

I Allelic imbalance, often differentialI Association with isoform usage (splicing QTLs)I RNA editing (allele absent from genome)

ChIP-seq: allele-specific binding

Variant calls are more general than genotypesGenotypes make additional assumptions

I A genotype identifies the set of alleles present at each locus.I The number of alleles (the ploidy) is decided and fixed.I Most genotyping algorithms output genotypes directly, under a

blind diploid assumption and special consideration of SNPsand haplotypes.

Those assumptions are not valid in general

I Non-genomic input (RNA-seq) does not represent a genotype.I Cancer genome samples are subject to:

I Copy number changesI Tumor heterogeneityI Tumor/normal contamination

So there is a mixture of potentially non-diploid genotypes, andthere is no interpretable genotype for the sample

Typical variant calling workflow

FASTQ

Typical variant calling workflow

AlignmentBWA GSNAP

FASTQ

gmapR

Typical variant calling workflow

AlignmentBWA GSNAP

FASTQ

FiltersRemove

PCR Dups(Picard)

Realign Indels

(GATK)

gmapR

Typical variant calling workflow

AlignmentBWA GSNAP

FASTQ

FiltersRemove

PCR Dups(Picard)

Realign Indels

(GATK)

Tallysamtools bam_tally gmapR

gmapR

ACT

TT A

C G

TTTTTTTTT

AA

A

T

TTTT

CCCCCCC

GGGGGGGGG

TTT

TTTTTTT

T

A CCCCCCCCC

GGGGGGGGGG

C

CCCC

GTTTC

CCC

A

AA

AAA

AA

A

AA

A

AA

AA

AC

Typical variant calling workflow

AlignmentBWA GSNAP

FASTQ

FiltersRemove

PCR Dups(Picard)

Realign Indels

(GATK)

Tallysamtools bam_tally gmapR

gmapR

CallingGATK VarScan2

VariantTools

ACT

TT A

C G

TTTTTTTTT

AA

A

T

TTTT

CCCCCCC

GGGGGGGGG

TTT

TTTTTTT

T

A CCCCCCCCC

GGGGGGGGGG

C

CCCC

GTTTC

CCC

A

AA

AAA

AA

A

AA

A

AA

AA

AC

ACT

TT A

C G

TTTTTTTTT

AA

A

T

TTTT

CCCCCCC

GGGGGGGGG

TTT

TTTTTTT

T

A CCCCCCCCC

GGGGGGGGGG

C

CCCC

GTTTC

CCC

A

AA

AAA

AA

A

AA

AA

AC

Het

Hom-alt

Low Freq

Error

POS REF ALT

2 T A4 A G8 C T

Typical variant calling workflow

AlignmentBWA GSNAP

FASTQ

FiltersRemove

PCR Dups(Picard)

Realign Indels

(GATK)

Tallysamtools bam_tally gmapR

gmapR

CallingGATK VarScan2

VariantTools

ACT

TT A

C G

TTTTTTTTT

AA

A

T

TTTT

CCCCCCC

GGGGGGGGG

TTT

TTTTTTT

T

A CCCCCCCCC

GGGGGGGGGG

C

CCCC

GTTTC

CCC

A

AA

AAA

AA

A

AA

A

AA

AA

AC

ACT

TT A

C G

TTTTTTTTT

AA

A

T

TTTT

CCCCCCC

GGGGGGGGG

TTT

TTTTTTT

T

A CCCCCCCCC

GGGGGGGGGG

C

CCCC

GTTTC

CCC

A

AA

AAA

AA

A

AA

AA

AC

Het

Hom-alt

Low Freq

Error

POS REF ALT

2 T A4 A G8 C T

Annotation Comparison

Sources of technical error

Errors can occur at each stage of data generation:I Library prepI SequencingI Alignment

Variant information for filtering

Information we know about each variant, and how it is useful:

Information UtilityBase Qualities Low quality indicates sequencing errorRead Positions Bias indicates mapping issuesGenomic Strand Bias indicates mapping issuesGenomic Position PCR dupes; self-chain, homopolymersMapping Info Aligner-dependent quality score/flags

Typical QC filters

10.1038/nbt.2514

These filters are heuristicsthat aim to reduce theFDR; however, they willalso generate falsenegatives and are bestapplied as soft filters(annotations).

Whole-genome sequencing and problematic regions

I Many genomic regions are inherently difficult to interpret.I Including homopolymers, simple repeats

I These will complicate the analysis with little compensatingbenefit and should usually be excluded.

Outline

Introduction

Calling variants vs. reference

Downstream of variant calling

VariantTools package

Visualization

VariantTools pipeline

At least twoalt reads

At least 4%alt read fractionC

all

Po

st F

ilter

Max Countin Neighborhood

Ou

tpu

t

Inp

ut

Variants

Tal

ly

Unique Alignments

MappingQuality > 13

Require > 23Base Quality

Mask SimpleRepeats

Ignore PicardDuplicates

QA

dbSNP positionsnot considered;

mostly useful for WGS

Not overlappingHP ( > 6nt )

Overlappingends in same pair

are clipped

Binomial LikelihoodRatio Test:

p(var) = 0.2 /p(error) = 0.001

UCSC self-chain as indicator of mappability

I UCSC publishes the self-chain score as a generic indicator ofintragenomic similarity that is independent of any aligner

I About 6% of the genome fits this definitionI Virtually all (GSNAP) multi-mapping is in self-chainsI Lower unique coverage in self-chains

Aligner matters: coverage and mappability

BWA coverage

GS

NA

Pco

vera

ge

0

50

100

150

200

0

50

100

150

200

0 50 100 150 200 0 50 100 150 200 0 50 100 150 200

FN

self−chained

FP

self−chained

TP

self−chained

unchained unchained unchained

FN FP TP

Aligning indels is error proneResolved by indel realignment

Homopolymers are problematic

Discard variants over or next to homopolymers (>6nt)

FAIL

PASS

CTGCGAAAAAAAA

CTGCGAAAAAAAA

0.0

0.1

0.2

0.3

inside/adjacent outside

Relationship to Nearest Homopolymer

FD

R

Choosing the homopolymer length cutoff

I We fit two logistic regressions to find the optimal length cutofffor our filter

I Response, TP: whether the variant call is a true positiveI Length as linear predictor:

I TP ~ I(hp.dtn <= 1) + hp.lengthI Indicator for when length exceeds 7:

I TP ~ I(hp.dtn <= 1) + I(hp.length > 7)

Logistic regression results

0.7

0.8

0.9

1.0

0.4 0.6 0.8 1.0Fraction of FP retained

Fra

ctio

n of

TP

ret

aine

d

group TP ~ I(dtn.hp <= 1) + hp.length TP ~ I(dtn.hp <= 1) + I(hp.length > 7)

sample 10 YRI x 90 CEU 50 YRI x 50 CEU 90 YRI x 10 CEU

Effect of coverage extremes on frequencies

0

2

4

6

8

0.00 0.25 0.50 0.75 1.00

altDepth/totalDepth

Den

sity

Coverage (1,40]

(40,120]

(120,Inf]I Coverage sweet-spot

(40-120) matches expecteddistribution.

I High coverage (>120) hasmuch lower frequencies thanexpected; mapping error?

I Low coverage also different

Coverage extremes and self-chained regions

self−chained unchained

0.0

0.1

0.2

0.3

0.4[0

,10]

[10,

20]

[20,

30]

[30,

40]

[40,

50]

[50,

60]

[60,

70]

[70,

80]

[80,

90]

[90,

100]

[100

,Inf]

[0,1

0]

[10,

20]

[20,

30]

[30,

40]

[40,

50]

[50,

60]

[60,

70]

[70,

80]

[80,

90]

[90,

100]

[100

,Inf]

Coverage Bin

FD

R

Variant density filter performance

Discard variants clumped on the chromosome.

FAIL

PASS

0.0

0.1

0.2

0.3

0.4

0.5

> 0.1 ≤ 0.1

Neighborhood Score

FD

R

Outline

Introduction

Calling variants vs. reference

Downstream of variant calling

VariantTools package

Visualization

Downstream of variant calling

Calling(vs reference)

GATK VarScan2

VariantTools

POS REF ALT

2 T A4 A G8 C T

Functional Annotations

Genomic context, coding consequences, disease assocations

Annovar Ensembl VEP

VariantAnnotation

VariantFiltering

Two sample comparisons

Mutation callingRNA-editing

ensemblVEP

mutect strelka

VariantTools

Interpretation

VarScan2

Direct

Calling mutations through filtering

I We have two sets of variant calls (vs. reference) and need todecide which are specific to one (i.e., the tumor)

I We have to decide whether the variant frequency is:I Non-zero in tumor butI Zero in normal

I Variant frequencies are a function of:I Copy number changesI Tumor/normal contaminationI Sub-clonality (tumor heterogeneity)I Mutations

I Mutations often present at low frequency and may even showup in the normal data due to contamination

VariantTools mutation calling algorithm

A mutation must pass the following filters:I The variant was only called in the tumorI There was sufficient coverage in normal to detect a variant,

assuming the likelihood ratio model and given a power cutoffI The raw frequency in normal is sufficiently lower than the

frequency in tumor (avoids near-misses in normal)

Functional annotations with VariantAnnotation

The VariantAnnotation package

I Handles import/export of variants from/to VCFI Defines central data structures for representing variants

I VCF objects represent full complexity of VCF as a derivative ofSummarizedExperiment

I VRanges extends GRanges for special handling of variantsI Annotates variants with:

I Genomic context: locateVariants()I Coding consequences: predictCoding()I SIFT/PolyPhen

I Filters VCF files as a stream (filterVcf())

Learn moreThursday lab on annotating variants

Outline

Introduction

Calling variants vs. reference

Downstream of variant calling

VariantTools package

Visualization

Overview

I Convenient interface for tallying mismatches and indelsI Several built-in variant filtersI Combines filters into a default calling algorithmI Other utilities: call wildtype, ID verificationI Integrates:

I VRanges data structure from VariantAnnotationI Tallying with bam_tally via gmapRI FilterRules framework from IRanges

Tallying

The underlying bam_tally from Tom Wu’s GSTRUCT accepts anumber of parameters, which we specify as a TallyVariantsParamobject. The genome is required; we also mask out the repeats.

library(VariantTools)data(repeats, package = "VariantToolsTutorial")param <- TallyVariantsParam(TP53Genome(), mask = repeats)

Tallies are generated via the tallyVariants function:tallies <- tallyVariants(bam, param)

VRanges

I The tally results are stored in a VRanges objectI Extension of GRanges to describe variantsI One element/row per position + alt combinationI Adds these fixed columns:

ref ref allelealt alt alleletotalDepth total read depthrefDepth ref allele read depthaltDepth alt allele read depthsampleNames sample identifierssoftFilterMatrix FilterMatrix of filter resultshardFilters FilterRules used to subset object

VRanges features

I Rough, lossy, two-way conversion between VCF and VRangesI Matching/set operations by position and alt (match, %in%)I Recurrence across samples (tabulate)I Provenance tracking of applied hard filtersI Convenient summaries of soft filter results (FilterMatrix)I Lift-over across genome builds (liftOver)I VRangesList, stackable into a VRanges by sampleI All of the features of GRanges (overlap, etc)

Tally statistics

In addition to the alleles and read depths, tallyVariants provides:

Raw counts Count before quality filter for alt/ref/totalMean quality Mean base quality for alt/refStrand counts Plus/minus counts for alt/refUniq read pos Number of unique read positions for alt/refMean read pos Mean read position (cycle) for alt/refVar read pos Variance in read position for alt/refMDFNE Median distance from nearest end for alt/refRead pos bins Counts in user-defined read pos bins for alt

Filtering framework

VariantTools implements its filters within the FilterRules frameworkfrom IRanges. The default variant calling filters are constructed byVariantCallingFilters:calling.filters <- VariantCallingFilters()

Post-filters are filters that attempt to remove anomalies from thecalled variants:post.filters <- VariantPostFilters()

Filter tallies into variant calls

The filters are then passed to the callVariants function:variants <- callVariants(tallies, calling.filters,

post.filters)

Or more simply in this case:variants <- callVariants(tallies)

Interoperability via VCF

We can export the variant calls to a VCF file:writeVcf(variants, "variants.vcf", index = TRUE)

Outline

Introduction

Calling variants vs. reference

Downstream of variant calling

VariantTools package

Visualization

Visualizing variants with IGV SRAdb

Creating a connection to IGV

library(SRAdb)startIGV("lm")sock <- IGVsocket()

Exporting our calls as VCF

vcf <- writeVcf(variants, "variants.vcf", index = TRUE)

Creating an IGV session

Create an IGV session with our VCF, BAMs and custom p53genome:rtracklayer::export(genome, "genome.fa")session <- IGVsession(c(bam.paths, vcf), "session.xml",

"genome.fa")

Load the session:IGVload(sock, session)

Browsing regions of interest

IGV will (manually) load BED files as a list of bookmarks:rtracklayer::export(interesting.variants, "bookmarks.bed")

IGV section, from R

VariantExplorer package

I The VariantExplorer package by Julian Gehring is anunreleased package for visually diagnosing variant calls

I Produces static ggbio plots and interactive web-based plotsbased on epivizr

I The epivizr package (Hector Corrada Bravo) is abrowser-based genomic visualization platform that pulls datadirectly from a running R session

I Get epivizr:devtools::install_github("epivizr", "epiviz")