Enhanced structural variant and breakpoint detection using SVMerge by integration of multiple detection methods and local assembly

Vertebrate Resequencing Informatics 22nd March, 2011

Enhanced structural variant and breakpoint detection using SVMerge by integration of multiple detection methods and local assembly

Kim Wong/Thomas Keane Vertebrate Resequencing Informatics

http://svmerge.sourceforge.net


Genomic Structural Variation

Large DNA rearrangements (>100bp) Frequent causes of disease  Referred to as genomic disorders  Mendelian diseases or complex traits such as behaviors

 E.g. increase in gene dosage due to increase in copy number  Prevalent in cancer genomes

Many types of genomic structural variation (SV)  Insertions, deletions, copy number changes, inversions,

translocations & complex events Comparative genomic hybridization (CGH) traditionally used to for copy number discovery  CNVs of 1–50 kb in size have been under-ascertained

Next-gen sequencing revolutionised field of SV discovery  Parallel sequencing of ends of large numbers of DNA fragments  Examine alignment distance of reads to discover presence of

genomic rearrangments  Resolution down to ~100bp


Simple types of Structural Variation


Deletion

SV Visualisation  LookSeq viewer  Read pairs displayed  Y axis is aligned insert size

Deletions are easily spotted  Read pairs are mapped

further apart than expected  Coverage is zero across

the deletion sequence Deletion in NOD/ShiLtJ


Inversion

Mate pairs align in the same orientation

Coverage zero at breakpoints


Insertion

One end mapped reads

Coverage zero at breakpoint


Insertion Insertion

Inversion

Complex SV Events


Human Examples

Stankiewicz and Lupski (2010) Ann. Rev. Med.


Example 2: Transposable element insertion in mice


SVMerge

Initially developed for mouse genomes project  Several software packages currently available to discover SVs

Various approaches using information from anomalously mapped read pairs OR read depth analysis No single SV caller is able to detect the full range of structural variants  Paired-end mapping information, for example, cannot detect SVs where the

read pairs do not flank the SV breakpoints   Insertion calls made using the split-mapping approach are also size-limited

because the whole insertion breakpoint must be contained within a read  Read-depth approaches can identify copy number changes without the need

for read-pair support, but cannot find copy number neutral events SVMerge, a meta SV calling pipeline, which makes SV predictions with a collection of SV callers   Input is a BAM file per sample  Run callers individually + outputs sanitized into standard BED format  SV calls merged, and computationally validated using local de novo assembly  Primarily a SV discovery/calling + validation tool


SVMerge Workflow

Wong et al (2010)


SV Callers

Wong et al (2010)


Local Assembly Validation

Key to the approach is the computational validation step  Local assembly and breakpoint refinement  All SV calls (except those lacking read

pair support e.g. CNG/CNL) Algorithm  Gather mapped reads, and any

unmapped mate-pairs (<1kb of a insertion breakpoint, <2kb of all other SV types)

 Run local velvet assembly  Realign the contigs produced with

exonerate  Detect contig breaks proximal to the

breakpoint(s)


Breakpoint Improvement (simulated)


Breakpoint Improvement (Real data)

Yalchin and Wong et al, in prep


Application to HapMap trio dataset

High-depth HapMap trio (NA18506, NA18507, NA18508)  42x, 42x and 40x

Reads processed through Vert. Reseq. Pipeline  Aligned to the GRCh37 human reference using BWA  Single BAM file for each individual

BreakDancerMax, Pindel, RDXplorer, SECluster, and RetroSeq Exclude calls  600 bp from a reference sequence gap  1 Mb from a centromere or telomere

Computational validation of raw candidate calls


NA18506 Results


Does multiple callers discover more SVs?


How do the calls measure up?

Compared the overlap of the deletion, gain, and inversion calls against the curated Database of Genomic Variants  Overlapped with calls in DGV at a rate significantly higher than

expected by random chance  Deletions in DGV: 71% (NA18506), 81% (NA18507), and 71%

(NA18508)  Copy number gains in DGV: 29% (NA18506), 32% (NA18507),

and 36% (NA18508)  Inversions in DGV: 47% (NA18506), 69% (NA18507), and 51%

(NA18508) Child calls not in DGV also called in the parents  Further 18% deletions, 32% inversions, 54% duplications  Estimated max. false positive rate of 11%, 21%, and 17%

All child-only SV calls comprise 11% of the child's final SV call  Considerable improvement from 'merged raw’ (50% unique)


Complex SV Types

Yalchin and Wong et al, in prep


Future Work

SVMerge primarily a discovery and validation tool  Extensible pipeline so that calls from any method to be easily

incorporated Developed primarily for mouse genomes project  Successfully applied to human trio dataset  Computationally validation approach reduces false positives

Complex SVs  Cataloging repeating combinations of multiple SV events in small

loci 2011 development  Low coverage cross-population SV discovery  Genotyping existing SVs in new samples  Better support for heterozygous calls  Integration of SVMerge into Vert. Reseq. pipeline for UK10K

Technology

Enhanced structural variant and breakpoint detection using SVMerge by integration of multiple detection methods and local assembly