methylation-aggregation

Aggregate DNA Methylation Data Across Studies

When to Use

• User wants to build a tissue-level DNA methylation landscape from multiple WGBS experiments
• User asks "where is DNA methylated in brain?" or "find hypomethylated regions across donors"
• User needs to identify HMRs (hypomethylated regions), UMRs, or PMDs from aggregated WGBS data
• User wants per-CpG weighted methylation averages from multiple experiments
• Example queries: "aggregate WGBS data for liver", "build methylation map across donors", "find unmethylated CpG islands in pancreas"

Build a comprehensive methylation landscape for a tissue/cell type by merging WGBS bedMethyl files from multiple ENCODE experiments.

Scientific Rationale

The question: "What is the DNA methylation state across the genome in my tissue?"

DNA methylation is fundamentally different from histone marks and accessibility:

| Property | Histone/Accessibility | DNA Methylation | |----------|----------------------|-----------------| | Signal type | Binary (bound/open or not) | Continuous (0-100% methylated) | | Default state | Unmarked | ~70-80% methylated (CpG context) | | Biology of interest | Where marks ARE present | Where methylation is ABSENT or REDUCED | | Aggregation approach | Union of peak calls | Average/median of methylation levels per CpG |

The key insight: Unlike histone ChIP-seq where we want the union of all peaks, for methylation we want the average methylation level per CpG site across individuals. Methylation is a quantitative, continuous signal measured at every CpG dinucleotide.

However, for identifying regulatory regions, we focus on hypomethylated regions (HMRs) — stretches of low methylation that mark active regulatory elements. HMRs can be treated more like peaks for union-style aggregation.

Literature Support

• Roadmap Epigenomics (Schultz et al. 2015, Nature, 2,900+ citations): Established that tissue-specific HMRs mark active regulatory elements; demonstrated per-CpG averaging across biological replicates as standard approach
• DMRcate (Peters et al. 2021, Nucleic Acids Research, 65 citations): Method for calling differentially methylated regions from multiple WGBS samples; uses kernel smoothing across CpG sites
• ENCODE Phase 3 (Gorkin et al. 2020, Nature, 301 citations): Integrated methylation data with histone marks and accessibility to define chromatin states
• ENCODE Blacklist (Amemiya et al. 2019, Scientific Reports, 1,372 citations): Problematic genomic regions to filter. DOI
• Zhou et al. 2020 (Nature Genetics): Tissue-specific methylation patterns: ~80% of CpGs are constitutively methylated, ~10% constitutively unmethylated (CpG islands/promoters), ~10% tissue-variable
• Liu et al. 2024 (Briefings in Bioinformatics): Cross-platform comparison (NovaSeq vs DNBSEQ) showing WGBS is gold standard; coverage depth critically affects accuracy; platform differences exist in GC-rich regions
• Ortega-Recalde et al. 2021 (Methods in Molecular Biology): Demonstrated that even low-coverage WGBS can accurately estimate global methylation levels, with bootstrap methods to quantify uncertainty

Two-Level Analysis

1. Per-CpG level: Average methylation at each CpG site across samples (quantitative map)
2. Region level: Identify HMRs, PMDs, and UMRs from the averaged profile (union of regulatory regions)

Step 1: Find All Available WGBS Data

encode_search_experiments(
    assay_title="WGBS",
    organ="pancreas",           # user's tissue of interest
    biosample_type="tissue",
    limit=100
)

Present a summary to the user:

• Total WGBS experiments
• Labs represented
• Unique donors/biosamples
• Genome coverage per experiment

Use encode_get_facets to check availability:

encode_get_facets(assay_title="WGBS", organ="pancreas")

Note: WGBS is expensive to generate. Typical tissues have 2-5 experiments. Even 2 biological replicates are valuable for identifying consistent methylation patterns.

Step 2: Quality-Gate Each Experiment

encode_get_experiment(accession="ENCSR...")

WGBS Quality Checks

• Audit status: no ERROR flags
• Bisulfite conversion rate: >=98% (measured by lambda spike-in or non-CpG methylation)
• Genome coverage: >=10x mean CpG coverage for reliable per-site estimates
• Mapping rate: >=50% (bisulfite-converted reads are harder to map)
• Duplication rate: <30%
• Has methylation state at CpG output files (bedMethyl format)

Include if:

• Bisulfite conversion >=98%
• Mean CpG coverage >=10x
• Has bedMethyl output files for GRCh38

Exclude if:

• ERROR audit flags
• Conversion rate <98% (unconverted reads create false methylation calls)
• Very low coverage (<5x mean) — individual CpG estimates unreliable

Track all included experiments:

encode_track_experiment(accession="ENCSR...")

Step 3: Download bedMethyl Files

For each experiment:

encode_list_files(
    experiment_accession="ENCSR...",
    output_type="methylation state at CpG",
    assembly="GRCh38"
)

bedMethyl format (ENCODE standard):

chr  start  end  name  score  strand  thickStart  thickEnd  color  coverage  percentMethylated

• Column 10: read coverage at this CpG
• Column 11: percent methylation (0-100)

Prefer preferred_default=True files:

encode_download_files(
    file_accessions=["ENCFF...", ...],
    download_dir="/path/to/data/methylation",
    organize_by="flat"
)

Step 4: Per-Sample Quality Filtering

4a. Coverage Filtering (CRITICAL)

Low-coverage CpGs have unreliable methylation estimates. Filter per-sample:

# Keep only CpGs with >= 5x coverage (column 10)
# More stringent: >= 10x for quantitative analysis
awk '$10 >= 5' sample.bedMethyl > sample.covfiltered.bedMethyl

Coverage thresholds by use case: | Threshold | Use Case | Typical CpGs Retained | |-----------|----------|----------------------| | >=3x | Exploratory / maximum retention | ~90% of CpGs | | >=5x | Standard analysis | ~80% of CpGs | | >=10x | High-confidence quantitative | ~60% of CpGs |

4b. ENCODE Blocklist Filtering (Amemiya et al. 2019)

# Download from: https://github.com/Boyle-Lab/Blacklist/blob/master/lists/hg38-blacklist.v2.bed.gz
bedtools intersect -a sample.covfiltered.bedMethyl -b hg38-blacklist.v2.bed -v > sample.filtered.bedMethyl

4c. Strand Merging (Optional but Recommended)

CpG methylation is typically symmetric (same on both strands). Merge strand-specific calls to increase per-CpG coverage. Caveat: Some ENCODE bedMethyl files may already be strand-merged — check if both strands are present before applying this step:

# Group CpGs by position (forward and reverse strand of same CpG)
# Sum coverage, calculate weighted average methylation
awk 'BEGIN{OFS="\t"} {
    # CpG position (use the C position as canonical)
    if ($6 == "+") pos = $2
    else pos = $2 - 1
    key = $1"\t"pos
    cov[key] += $10
    meth[key] += ($11/100) * $10
} END {
    for (k in cov) {
        split(k, a, "\t")
        avg_meth = (meth[k] / cov[k]) * 100
        print a[1], a[2], a[2]+2, "CpG", 0, ".", a[2], a[2]+2, "0,0,0", cov[k], avg_meth
    }
}' sample.filtered.bedMethyl | sort -k1,1 -k2,2n > sample.merged_strands.bedMethyl

Step 5: Cross-Sample Aggregation (Per-CpG Averaging)

5a. Create a Unified CpG Matrix

# Step 1: Find CpGs covered in at least M of N samples
# Extract positions from each sample
for f in sample*.merged_strands.bedMethyl; do
    awk 'BEGIN{OFS="\t"} {print $1, $2, $3}' "$f"
done | sort -k1,1 -k2,2n | uniq -c | \
awk -v M=2 '$1 >= M {print $2, $3, $4}' OFS="\t" > shared_cpgs.bed

# Step 2: For each sample, extract methylation at shared CpGs
for f in sample*.merged_strands.bedMethyl; do
    bedtools intersect -a shared_cpgs.bed -b "$f" -wa -wb | \
    awk 'BEGIN{OFS="\t"} {print $1, $2, $3, $NF, $(NF-1)}' > "${f%.bedMethyl}.shar