Higher-order functional domains in the human ENCODE regions

This webpage is intended as a supplement to the manuscript, "Higher-order functional domains in the human ENCODE regions," by R.E. Thurman, N. Day, W.S. Noble, and J.A. Stamatoyannopoulos, Genome Research, 2007, 17, 991-994. Below find data, links to data, and analysis results referenced in that paper. There are also some results here included in the ENCODE Nature paper, "Identification and analysis of functional elements in 1% of the human genome by the ENCODE pilot project", Nature, 2007, 447, 799-816 (see 6-track segmentation below).

---

As part of the ENCODE project we have access to a number of datatypes, representing a variety of functional variables (e.g., DNaseI sensitivity, transcription, histone modifications, conservation, etc.) sampled in a nearly continuous fashion across the genome. We would like to understand to what degree such data reveal coherent higher-order features that may in turn illuminate the underlying functional architecture of the genome. To address this, we aim to develop approaches based on wavelet analysis for the discovery of "domain-level" behavior in fine scale data, and for correlating these apparently disparate functional data types.

---

The rough outline of methods used in the paper are as follows:

(postscript version of this diagram)

• Use wavelets and HMMs to define functional domains based on one or more continuous functional datasets from the ENCODE project. Those datasets, and their public availability, are listed below. Datasets available from the UCSC Genome Browser were downloaded under the ENCODE section using hg17 (May, 2004 assembly) coordinates.
  • Histone modifications
  • H3ac, H4ac, H3K4me1, H3K4me2, H3K4me3 available from the Sanger institute, UCSC Genome Browser tracks encodeSangerChipH3acHela, encodeSangerChipH4acHela, encodeSangerChipH3K4me1Hela, encodeSangerChipH3K4me2Hela, and encodeSangerChipH3K4me3Hela.
  • H3K27me3 available from the University of San Diego, UCSC Genome Browser track encodeUcsdChipH3K27me3.
  • RNA expression levels from Affymetrix, UCSC Genome Browser track encodeAffyRnaHeLaSignal
  • DNA replication timing from the University of Virgina, UCSC Genome Browser track encodeUvaDnaRepTr50
  • density of conserved non-coding sequence elements, from the ENCODE MSA (multi-species alignment) analysis subgroup. This data to appear on the UCSC Genome Browser. Links to the actual data used for the paper are here:
    • MSA "moderate" non-exonic raw data (coordinates of non-exonic conserved elements defined using the MSA group's "moderate" criteria).
    • Density of above elements in 3kb windows stepping at 1kb intervals
• "Validate" these domains in a variety of ways. For instance, exhibit statistically significant differences between HMM states in
  • percent coding sequence
  • # tx start sites
  • # CNS
  • MSA moderate non-exonic (number of these bases in a sliding window)
  • % CpG content

---

Software

• For wavelet smoothing using the maximal-overlap discrete wavelet transform (MODWT), we use R package waveslim.
• For HMM segmentation we use our own software package, HMMSeg (manuscript "Unsupervised segmentation of continuous genomic data," currently in review).
• Utility routines in R for data pre-processing (interpolation using adaptive loess fitting is included in R function UCSC2bed in file bed.R.
  • bed.R
  • interp.R

---

Annotation results

Segmentations

• Single-track segmentations

1. Sanger Chip (HeLa), 64kb wavelet smooth
• H3K4me1
• H3K4me2
• H3K4me3
• H3ac
• H4ac
2. Affy RNA Signal, 51.2kb wavelet smooth
3. UCSD Chip, H3K27me3, 64kb wavelet smooth
4. UVa DNA Replication, TR50
5. MSA non-exonic, 3kb sliding window, 64kb wavelet smooth

• Multiple-track segmentations

1. 3 track: H3ac, H3K27me3, and TR50 (HeLa), wavelet-smoothed as above
• Coordinates of domain segments in BED format (NOTE: "0" = active state, "1" = inactive state)
2. 4 track: Affy RNA, H3ac, H3K27me3, and TR50 (HeLa), wavelet-smoothed as above
• Coordinates of domain segments in BED format (NOTE: "1" = active state, "-1" = inactive state)
3. 5 track: Affy RNA, H3ac, H3K27me3, MSA non-exonic, and TR50 (HeLa), wavelet-smoothed as above
• Coordinates of domain segments in BED format (NOTE: "1" = active state, "-1" = inactive state)
4. 6 track: Affy RNA, H3ac, H3K27me3, TR50, merged HS density, RFBR density (HeLa) -- for ENCODE Nature paper
• Coordinates of domain segments in BED format (NOTE: "1" = active state, "-1" = inactive state)

Segmentation comparisons

Concordance figures are the percentage of bases whose state assignments agree in both segmentations.

1. Single-track comparisons

	Affy RNA Signal	UCSD Chip H3K27me3	MSA non-exonic	Uva TR50
Sanger H3K4me1	67% concordance	62% concordance	56% concordance	74% concordance
Sanger H3K4me2	72% concordance	55% concordance	58% concordance	69% concordance
Sanger H3K4me3	68% concordance	52% concordance	53% concordance	60% concordance
Sanger H3ac	75% concordance	59% concordance	63% concordance	70% concordance
Sanger H4ac	74% concordance	63% concordance	61% concordance	75% concordance
Affy RNA Signal		49% concordance	57% concordance	61% concordance
UCSD Chip H3K27me3			54% concordance	70% concordance
MSA non-exonic				57% concordance

2. Sanger histone modifications vs. themselves

	H3ac	H4ac	H3K4me1	H3K4me2	H3K4me3
H3ac
H4ac	85% concordance
H3K4me1	73% concordance	80% concordance
H3K4me2	86% concordance	80% concordance	77% concordance
H3K4me3	80% concordance	73% concordance	68% concordance	87% concordance

3. Tissue differences: HeLa vs. GM

1. Sanger H3K4me2, 64kb smooth 76% concordance
2. Sanger H3K4me1, 64kb smooth 69% concordance
3. Sanger H3K4me3, 64kb smooth 78% concordance
4. Sanger H3ac, 64kb smooth 72% concordance
5. Sanger H4ac, 64kb smooth 76% concordance
6. Affy RNA Signal, 51.2kb smooth 81% concordance

4. Wavelet scale differences

1. SangerH3K4me2, GM, 2kb smooth vs. 64kb smooth 70% concordance
2. SangerH3K4me2, segment size as a function of scale (postscript file)
3. Segment boundary scale sensitivity, based on 4-track (Affy zero-threshold, H3ac, H3K27me3, and TR50) segmentations at 32kb, 64kb, and 128kb scales.

5. Affy/H3ac/H3K27me3/TR50 wavelet 4-track vs. each of its component single-track results

1. Sanger H3ac 64kb (89% concordance)
2. Affy RNA, 51.2kb (80% concordance)
3. UCSD Chip, H3K27me3 64kb (62% concordance)
4. TR50 (76% concordance).

6. Affy/H3ac/H3K27me3/MSA non-exonic/TR50 wavelet 5-track vs. each of its component single-track results

1. Sanger H3ac 64kb (90% concordance).
2. Affy RNA, 51.2kb (79% concordance).
3. UCSD Chip, H3K27me3 64kb (62% concordance).
4. TR50 (76% concordance).
5. MSA non-exonic (62% concordance)

7. 3-track vs. 4-track

1. H3ac/H3K27me3/Tr50 vs Affy/H3ac/H3K27me3/MSA/Tr50, 78% concordance

8. 4-track vs. 4-track

1. Affy/H3ac/H3K27me3/Tr50 vs. H3ac/H3K27me3/MSA/Tr50, 78% concordance

9. 4-track vs. 5-track

1. Affy/H3ac/H3K27me3/Tr50 vs Affy/H3ac/H3K27me3/MSA/Tr50 (wavelet-smoothed, zero Affy threshold) 98% concordance

High-confidence segments

1. 4 track: Affy (zero threshold), H3ac, H3K27me3, TR50 (wavelet)
2. 5 track: Affy (zero threshold), H3ac, H3K27me3, MSA non-exonic, TR50 (wavelet)
3. Affy, H3ac, H3K27me3, TR50 - single-track segmentations only
• Compare high-confidence regions with multi-track segmentation

---