Identifying Structural Variations Using Next Generation Sequencing Data

1
Identifying Structural Variations Using Next
Generation Sequencing Data

• Seunghak Lee, Elango Cheran, Michael Brudno
• Department of Computer Science
• University of Toronto
• ISMB 2008 SIG on NGS technologies

2
Computational Methods to Detect Structural
Variants

• Direct comparison of genomes (Levy et al. 2007)
• No limitation of resolutions
• Expensive to assemble the whole genome
• Unassembled clone-end data (Tuzun et al. 2005,
  Korbel et al. 2007)
• Use of matepairs, part of data for assembling
  whole genome
• Unable to exploit high coverage of reads
• Probabilistic framework for clone-end data (Lee
  et al. 2008)
• Based on unassembled clone-end data
• Takes advantage of high coverage of reads

3
Outline

• Probabilistic framework
• New Measure for Clusters
• Results

4
Outline

• Probabilistic framework
• New Measure for Clusters
• Results

5
What are Matepairs?
DNA fragment
ATCAA
CTAAG
Insert size
6
Overview of Probabilistic Framework

• We defined pairwise probabilities (Lee et al.
  2008)
• P(Xi, Xjins), P(Xi, Xjdel), P(Xi, Xjinv),
  P(Xi, Xjtrans)

Xj
Xi
Donor
Xj
Xi
REF
7
Overview of Probabilistic Framework

• We defined pairwise probabilities (Lee et al.
  2008)
• P(Xi, Xjins), P(Xi, Xjdel), P(Xi, Xjinv),
  P(Xi, Xjtrans)

Xj
Xi
Donor
Xj
Xi
REF
8
Probabilistic Framework - Insertion
?
Xj
Xi
Donor
Xj
Xi
REF
9
Probabilistic Framework - Insertion
Xj
Xi
Donor
Xj
Xi
REF
10
Probabilistic Framework - Insertion

• Let si mapped distance of Xi in REF
• mi insert size of Xi

Estimated size of insertion explained by Xi is ri
si-mi
Xj
Xi
Donor
mi
Xj
Xi
REF
si
11
Probabilistic Framework - Insertion

• Size of insertion

Xj
Xi
r
Donor
mj
mi
Xj
Xi
REF
sj
si
12
Probabilistic Framework - Insertion

• si r average insert size of Xi

Xj
Xi
r
Donor
Insert size of Xi
mi
Xj
Xi
REF
si
13
Probabilistic Framework - Insertion

• Now, we define with

Y random variable for insert size
p(Y)
(si r) is close to µ gt more likely insertion
si r
Average insert size of Xi
14
Probabilistic Framework - Insertion
p(Y)
(si r) is close to µ gt more likely insertion
si r
Average insert size of Xi
15
Probabilistic Framework - Insertion

• Shaded area

p(Y)
(si r) is close to µ gt more likely insertion
µ - (si r)
si r
Average insert size of Xi
16
Outline

• Probabilistic framework
• New Measure for Clusters
• Results

17
Motivation

• Probabilities were defined by pairwise
  comparisons
• High coverage of matepairs are now available from
  NGS technologies
• Expect large number of matepairs in each cluster
• We developed an alternative measure which takes
  into account all matepairs in each cluster
  together

18
Kullback-Leibler Divergence Overview

• KL divergence is a measure of the difference of
  two probability distributions
• KL divergence of Q from P
• P true distribution
• Q model distribution

19
A New Measure for Clusters (1)
p(Y)

• p(Y) observed distribution of an insert size
• q(Y) shifted distribution of p(Y) with mean of
  sr, where s is mapped distance and r is size of
  insertion

(s r) is close to µ gt more likely insertion
KL(pq) 0
µ
20
A New Measure for Clusters (2)

• Define a new measure using KL divergence taking
  into account of all matepairs in cluster Ck
• p(i)(Y) observed distribution of insert size of
  i-th matepair in Ck
• q(i)(Y) shifted distribution of p(i)(Y) with
  mean of s(i)r

21
Optimizing New Measure

• Assuming p(i)(Y) is a Gaussian with given mean
  and variance
• Find optimum r by minimizing DKL(Ck) using
  gradient descent optimization method
• Optimized DKL(Ck) is used as an alternative
  measure for a cluster Ck

22
Use of KL Divergence for Clustering

• We cluster mappings of matepairs using
  Hierarchical Clustering
• Linkage distance between Cu and Cv
• Find two closest clusters if DKL(Cu,Cv)lt
  cutoff, then merge

Cv
Ck
Cu
REF
23
Outline

• Probabilistic Framework
• New Measure for Clusters
• Results

24
Clustering Results

• We started with 170,000,000 matepairs
• 34.8 were uniquely mapped (93 for Sanger
  reads)
• 97.5 had a concordant position
• Through the clustering procedure we found (FDR
  0.05)
• 152 Insertion clusters (4 had a uniquely mapped
  matepair)
• 930 Deletion clusters (43)
• 55 Inversion clusters (1)
• 19 Translocation (cross-chromosome) cluster
  (all were required to have a uniquely mapped read)

25
Changes in Coverage of Deletion Clusters

• Coverage of deletions
• Homozygous No reads at locus of deletion
• Heterozygous Half reads at locus of deletion

Chr1
Chr1
26
Example of Deletion Cluster

• A deletion cluster in chr11,213,000-1,215,000
• Contains 16 discordant matepairs
• Overlaps deletion in Levy
• Size of deletion 2,000bp

27
Distribution of Coverage of Deletions (1)

• r1r2r3
• C(ri) is the number of reads in region ri
• If this is a real deletion, coverage in r2 should
  be less than average coverage of r1 and r3
• We computed fraction of coverage of deletion to
  neighbors

deletion
REF
r2
r3
r1
F0 (Homozygous) F0.5 (Heterozygous)
28
Distribution of Coverage of Deletions (2)
Number of clusters
F

• 699/930 (75) deletions have less than 0.7
  coverage of reads in neighbor regions (Flt0.7)

29
Agreement with Previous Results

• All of the correlations are significant
  (p-values lt 0.001 via Monte Carlo)
• Found 444 novel structural variants

30
Conclusions

• Used a probabilistic framework for finding
  structural variants with NGS data
• Isolated insertions, deletions, inversions and
  translocations between the reference public human
  genome and AB-SOLiD data
• 75 of deletions supported by low read coverage
• Strong correlation between our results and
  previous studies

31
Acknowledgments

• Michael Brudno
• Elango Cheran
• Francisco de la Vega
• Andy Pang
• Lars Feuk
• Stephen Scherer
• Rest of UofT DSC CompBio Group

Applied Biosystems
Sick Kids Hospital