Multiscale correlations in continuous genomic data

1
Multi-scale correlations in continuous genomic
data

• Bob Thurman, Ph.D.
• Research Scientist
• Stamatoyannopoulos and Noble Labs
• Department of Genome Sciences
• University of Washington
• 6 January, 2008
• PSB 2008, The Big Island of Hawaii

2
Wavelet correlations

• A technique using wavelets to uncover
  correlations between continuous genomic datasets
  at multiple scales.
• Two case studies from ENCODE project. Multiple
  measures of functionality. To what extent do
  they agree or disagree?
• DNaseI sensitivity (chromatin accessibility) vs.
  histone modifications.
• H3K4me2 (activating) vs. H3K27me3 (repressive)

3
Wavelet representation of continuous data
DNaseI
H3K4me2
Wavelet coefficient measures the strength of
the change of the signal at the given position,
when considered at the given scale. A tool for
time-frequency analysis.
4
Significance of correlations
5
Statistical significance via sampling
0/1000 samples with that enrichment (plt.001)
Sampling strategy based on correlations between
large chunks randomly sampled separately from
each dataset.
6
Sample size matters
For 500kb region, p 0.0004 (2/5000 exceeding
samples)
7
High correlation correlates with gene density
Each point corresponds to one of the 31 ENCODE
regions of size 500kb
8
Case 2 H3K27me3 vs. H3K4me2
2kb
scale
32kb
9
Significance of pattern of positive and negative
correlations
plt0.001 for the observed fraction of correlation
values above 0.5 and below -0.5
10
Use 2-state HMMs to identify hi/hi regions
11
H3K27me3/H3K4me2 hi/hi regions

• 329 segments, covering 3425074 total bp, or
  approximately 10 of ENCODE.
• Over-represented GO categories
• Six transcription-related terms
• Regulation of cellular, physiological and
  biological processes
• Development

12
Summary

• Tool for exploring correlations between any two
  continuous genomic datasets
• Multiple scales explored simultaneously
• Correlation values at each position enable local,
  regional and global analyses.

13
Acknowledgments

• ENCODE project for funding and data
• John Stam, Bill Noble, and all my homies in the
  Noble and Stam labs
• Don Percival (UW) for help with wavelets

14
Scale matters
Related recent work successful partition of
Barski modifications at domain scale.
41kb scale
Input resolution