compare-biosamples

Compare ENCODE Data Across Biosamples

When to Use

• User wants to compare ENCODE experiments across different tissues, cell lines, or biosamples
• User asks about "tissue comparison", "cell-type differences", "tissue-specific enhancers", or "cross-tissue"
• User needs to identify constitutive vs tissue-specific regulatory elements
• User wants to map data availability across multiple biosamples before integrative analysis
• Example queries: "compare H3K27ac between liver and pancreas", "what marks are tissue-specific?", "find constitutive promoters across all tissues"

Help the user systematically compare data availability and experiments across different biosamples to identify tissue-specific regulatory patterns, constitutive elements, and cross-tissue differences.

Scientific Rationale

Cross-biosample comparison is the foundation of understanding tissue-specific gene regulation. Regulatory elements -- particularly enhancers -- are the primary drivers of cell-type identity, with promoters being largely shared across tissues. Comparing the same assay across multiple biosamples reveals which regulatory elements are constitutive (shared) versus tissue-specific (unique to one or few cell types).

The core question: "Which regulatory features distinguish tissue A from tissue B, and which are shared?"

This requires careful matching of datasets, awareness of batch effects, and understanding of the biosample hierarchy to avoid confounding biological signal with technical variation.

Literature Foundation

#	Reference	Key Contribution
1	Roadmap Epigenomics Consortium 2015, Nature, DOI:10.1038/nature14248 (~5,810 cit)	Generated 111 reference epigenomes across tissues/cell types; established the framework for cross-tissue epigenomic comparison. Showed that enhancer chromatin states are the most tissue-variable elements.
2	ENCODE Phase 3 2020, Nature, DOI:10.1038/s41586-020-2493-4 (~1,656 cit)	Expanded functional annotations to 1.3M candidate cis-regulatory elements (cCREs) across hundreds of biosamples; defined tissue-activity indices for regulatory elements.
3	Andersson et al. 2014, Nature, DOI:10.1038/nature12787 (~1,500 cit)	FANTOM5 atlas of active enhancers across 808 samples; demonstrated that only ~5% of enhancers are active across all tissues, with the majority being highly tissue-specific.
4	Heintzman et al. 2009, Nature, DOI:10.1038/nature07917 (~2,200 cit)	Showed histone modifications distinguish cell types: H3K4me1/H3K27ac at enhancers are the most discriminating tissue-specific marks, while H3K4me3 at promoters is largely shared.
5	Thurman et al. 2012, Nature, DOI:10.1038/nature11232 (~2,000 cit)	Mapped accessible chromatin across 125 cell types; demonstrated that DNase I hypersensitive sites define cell-type identity and that accessibility patterns cluster by tissue of origin.
6	Leek et al. 2010, Nat Rev Genet, DOI:10.1038/nrg2825 (~1,200 cit)	Comprehensive review of batch effects in genomic data; showed that lab, platform, and processing date can dominate biological variation if not properly controlled.
7	Forrest et al. 2014, Nature, DOI:10.1038/nature13182 (~1,100 cit)	FANTOM5 promoter-level expression atlas across 975 samples; demonstrated that promoter usage (not just gene expression) is tissue-specific and defines cell identity.

Tissue-Specific Regulation Principles

Understanding what varies across tissues and what does not is essential before designing a comparison.

What Is Shared vs Tissue-Specific (Heintzman 2009; Andersson 2014)

Feature	Cross-Tissue Behavior	Implication for Comparison
Promoters (H3K4me3)	Largely shared (~70% active in most tissues)	Poor discriminators between tissues
Enhancers (H3K27ac + H3K4me1)	Highly tissue-specific (~5% shared across all tissues)	Best discriminators; focus comparison here
Chromatin accessibility (ATAC/DNase)	Moderate tissue-specificity (~20-30% shared)	Good secondary discriminator; clusters by tissue of origin
Polycomb repression (H3K27me3)	Tissue-specific (marks silenced developmental genes)	Useful for identifying repressed lineage programs
Gene expression (RNA-seq)	Moderate tissue-specificity	Housekeeping genes shared; tissue-specific TFs are key
CTCF binding	Largely constitutive (~70% conserved)	Defines structural boundaries; less tissue-variable
DNA methylation	Bimodal; enhancers show tissue-variable methylation	Hypomethylation at active enhancers is tissue-specific

Key Insight

H3K27ac at enhancers is the single most informative mark for distinguishing tissues (Heintzman et al. 2009, Roadmap 2015). If the user can only compare one mark across tissues, H3K27ac should be the first choice, followed by chromatin accessibility (ATAC-seq or DNase-seq).

ENCODE Biosample Hierarchy

Level	Description	Biological Relevance	Reproducibility	Caveats
Tissue	Primary tissue from donor (e.g., pancreas, liver)	Highest -- in vivo biology preserved	Lower -- donor variation, cell-type heterogeneity	Mixed cell populations; composition varies by donor age/sex/health
Primary cell	Cells isolated from tissue (e.g., hepatocytes, islets)	High -- enriched for cell type	Moderate -- isolation stress, limited passages	Isolation method alters phenotype; culture conditions matter
Cell line	Immortalized cells (e.g., K562, HepG2, GM12878)	Lower -- transformed phenotype	Highest -- clonal, reproducible	May not represent normal tissue biology; passage number matters
In vitro differentiated	Cells derived from stem cells (e.g., iPSC-derived cardiomyocytes)	Moderate -- model system	Moderate -- protocol-dependent	Differentiation efficiency varies; often immature phenotype
Organoid	3D self-organizing structures	Moderate-high -- recapitulates tissue architecture	Lower -- heterogeneous	Emerging data type in ENCODE; limited coverage

Tier 1 Cell Lines (Most Comprehensive ENCODE Data)

Cell Line	Origin	Cancer/Normal	Best For
K562	Chronic myelogenous leukemia	Cancer	Hematopoietic chromatin, TF binding, 3D genome
GM12878	Lymphoblastoid (EBV-transformed B cells)	Transformed-normal	Immune regulation, 3D genome (Rao et al. 2014 Hi-C reference)
H1-hESC	Human embryonic stem cells	Normal	Developmental regulation, bivalent chromatin

These three cell lines have the most complete multi-omic profiling in ENCODE. They are excellent positive controls for verifying comparison pipelines before applying to user-specific tissues.

Biosample Comparability Rules

• Same biosample type preferred: Compare tissue-to-tissue, cell line-to-cell line
• Cross-type comparisons require caution: Cell line vs tissue introduces both biological and technical confounders
• Donor matching: When comparing tissues, match for life_stage, sex, and age when possible
• Passage number matters for cell lines: Different passages of the same cell line can diverge epigenomically

Step 1: Define the Comparison Design

Clarify the comparison type with the user. Each design has different requirements:

Comparison Design Patterns

Design	Description	Required Matching	Key Tools	Best File Types
Cross-tissue (same assay)	Same mark/assay in different organs	Same assay, same target, same assembly, same biosample_type	encode_search_experiments, encode_get_facets	IDR thresholded peaks, fold change over control
Multi-omic (same tissue)	Multiple assays in one biosample	Same biosample_term_name, same assembly	encode_get_facets, encode_search_experiments	Depends on assay
Disease vs normal	Pathological vs healthy tissue	Same organ, same assay, matched demographics	encode_search_experiments with biosample filter	IDR thresholded peaks, gene quantifications
Developmental time course	Same tissue at different life stages	Same organ, same assay, different life_stage	encode_search_experiments with life_stage filter	Signal tracks, gene quantifications
Cell line vs primary tissue	Transformed vs in vivo	Same organ of origin, same assay	encode_search_experiments, encode_compare_experiments	IDR thresholded peaks
Cross-species	Human vs mouse homologous tissues	Same organ, same assay, different organism	encode_search_experiments with organism filter	Requires liftOver; use signal tracks

Ask the user: "What tissues/cell types are you comparing, and what assay are you focusing on?"

Step 2: Map Data Availability Across Biosamples

Use encode_get_facets to build an availability matrix before searching for specific experiments.

2a. Check What Exists for Each Biosample

# For each tissue of interest, discover available assays and targets
encode_get_facets(organ="pancreas")
encode_get_facets(organ="liver")
encode_get_facets(organ="brain")

2b. Check What Exists for a Specific Assay Across Tissues

# See which organs have Histone ChIP-seq data
encode_get_facets(assay_title="Histone ChIP-seq")

# See which organs have ATAC-seq data
encode_get_facets(assay_title="ATAC-seq")

2c. Build the Availability Matrix

Present to the user a matrix like:

Assay / Target	Pancreas tissue	Liver tissue	Brain tissue	K562	GM12878
H3K27ac ChIP-seq	3 exp	5 exp	8 exp	12 exp	10 exp
H3K4me3 ChIP-seq	2 exp	4 exp	6 exp	11 exp	9 exp
ATAC-seq	1 exp	3 exp	5 exp	4 exp	3 exp
RNA-seq	4 exp	6 exp	10 exp	15 exp	8 exp
WGBS	0	2 exp	3 exp	2 exp	2 exp

Highlight gaps: "Pancreas has no WGBS data -- comparison of methylation patterns will be limited to liver and brain."

Step 3: Identify Matched Datasets

For a valid cross-tissue comparison, datasets must be matched on technical parameters. Search each tissue:

encode_search_experiments(
    assay_title="Histone ChIP-seq",
    target="H3K27ac",
    organ="pancreas",
    biosample_type="tissue",
    limit=50
)

Matching Criteria Checklist

For each experiment pair across tissues, verify:

Parameter	Must Match?	How to Check
Assay title	Yes	Search filter
Target (for ChIP)	Yes	Search filter
Genome assembly	Yes	File metadata; use GRCh38
Biosample type	Recommended	Search filter
Organism	Yes	Search filter
Life stage	Recommended	Experiment metadata
Sex	Preferred	Experiment metadata
Pipeline version	Preferred	encode_get_experiment
Sequencing depth	Comparable (within 2x)	File metadata
Read length	Preferred	File metadata

Step 4: Check Pairwise Compatibility

Track candidate experiments and use encode_compare_experiments for each cross-tissue pair:

# Track experiments from each tissue
encode_track_experiment(accession="ENCSR_pancreas")
encode_track_experiment(accession="ENCSR_liver")

# Check compatibility
encode_compare_experiments(
    accession1="ENCSR_pancreas",
    accession2="ENCSR_liver"
)

The compatibility tool checks:

• Organism match
• Assembly match
• Assay type match
• Biosample differences (expected for cross-tissue comparison)
• Target match
• Replication strategy
• Lab differences (potential batch effect source)

For cross-tissue comparison: Biosample mismatch is expected -- it is the variable of interest. Focus on ensuring all other parameters match.

Step 5: Assess and Control for Batch Effects (Leek et al. 2010)

Batch effects are the most common confounder in cross-biosample comparisons. When experiments come from different labs, platforms, or processing dates, technical variation can dominate biological signal.

Known Batch Effect Sources

Source	Impact	Detection Method
Lab of origin	High -- different protocols, antibodies, cell handling	Check lab field; PCA of signal should not cluster by lab
Sequencing platform	Moderate -- read quality, GC bias	Check platform in file metadata
Library preparation date	Moderate -- reagent lots, operator variation	Check experiment date_released
Antibody lot	High for ChIP-seq -- different enrichment profiles	Check antibody_lot_reviews in experiment metadata
Pipeline version	Low-moderate -- different peak calling parameters	Check analysis pipeline version
Read length	Low-moderate -- affects mappability	Check read_length in file metadata

How to Detect Batch Effects

1. PCA of signal: If the first principal component of signal tracks separates by lab (not by tissue), batch effects dominate
2. Check housekeeping loci: H3K4me3 at housekeeping promoters should be consistent across tissues. If it varies, suspect technical confounders.
3. CTCF as control: CTCF binding is ~70% constitutive. If CTCF shows tissue-specific differences that correlate with lab, suspect batch effects.

How to Mitigate

• Use ENCODE uniform pipeline outputs: All experiments processed through the same pipeline are more comparable than custom-processed data
• Match pipeline versions: Use files processed by the same pipeline version when possible
• Use fold-change-over-control signal tracks: Normalized to input, reducing depth and background differences
• Document all technical differences: In the comparison metadata, record every known technical difference between datasets

Step 6: Select Comparable Files

For each matched experiment, select files that are directly comparable:

# Get the recommended files for each experiment
encode_list_files(
    experiment_accession="ENCSR...",
    preferred_default=True,
    assembly="GRCh38"
)

ENCODE Blacklist filtering (required before comparison): Before any cross-tissue comparison, remove peaks and signal in ENCODE Blacklist regions (Amemiya et al. 2019, Scientific Reports, 1,372 citations). Blacklisted regions produce artifactual signal that appears consistent across tissues, inflating the count of "constitutive" elements. They can also show variable signal due to copy number differences between cell lines, creating false tissue-specific hits. Filter before comparison:

• Human GRCh38: hg38-blacklist.v2.bed.gz from Boyle-Lab/Blacklist
• Mouse mm10: mm10-blacklist.v2.bed.gz
• Filter with: bedtools intersect -v -a peaks.bed -b blacklist.bed > peaks.filtered.bed

File Selection by Comparison Goal

Comparison Goal	File Type	Output Type	Why
Peak overlap / tissue-specific peaks	bed narrowPeak	IDR thresholded peaks	Binary: present or absent in each tissue
Quantitative signal comparison	bigWig	fold change over control	Normalized signal; comparable across experiments
Differential expression	tsv	gene quantifications	TPM/FPKM for cross-tissue expression comparison
Chromatin state annotation	bed narrowPeak	All histone marks	Required for ChromHMM/chromatin state analysis
Visualization / heatmaps	bigWig	signal of unique reads	Raw signal for deepTools or genome browser

Critical: Use Same Output Type Across All Tissues

Do NOT mix IDR thresholded peaks from one tissue with pseudoreplicated peaks from another. This introduces systematic differences in peak number and stringency that confound biological comparison.

Step 7: Build the Comparison Matrix

Assemble a structured metadata table for all experiments in the comparison:

| Biosample | Organ | Type | Assay | Target | Accession | Audit | Depth | Lab | Pipeline |
|-----------|-------|------|-------|--------|-----------|-------|-------|-----|----------|
| Pancreas  | pancreas | tissue | Histone ChIP | H3K27ac | ENCSR... | clean | 22M | Bernstein | v2.1 |
| Liver     | liver | tissue | Histone ChIP | H3K27ac | ENCSR... | warn  | 18M | Snyder | v2.1 |
| Brain     | brain | tissue | Histone ChIP | H3K27ac | ENCSR... | clean | 25M | Bernstein | v2.1 |

Use encode_summarize_collection after tracking all experiments for a bird's-eye view:

# After tracking all experiments
encode_summarize_collection()

Flag potential issues in the matrix:

• Depth differences >2x between tissues
• Different labs (batch effect risk)
• Different pipeline versions
• WARNING or ERROR audit flags
• Missing data for some tissues (incomplete comparison)

Step 8: Suggest Analysis Strategies by Comparison Type

Tissue-Specific Elements

Peaks (or signals) present in one tissue but absent in others. The canonical approach:

1. Download IDR thresholded peaks for each tissue
2. Use bedtools intersect -v to find peaks unique to each tissue
3. Annotate tissue-specific peaks with genomic features (promoter, enhancer, intergenic)
4. Expect: enhancer peaks will be the most tissue-variable (Heintzman 2009, Andersson 2014)

Constitutive Elements

Peaks present across ALL tissues compared:

1. Use bedtools multiintersect across all tissue peak files
2. Filter for regions present in all (or N-1) tissues
3. Expect: promoters and CTCF sites will dominate constitutive elements
4. Use as positive controls to verify comparison pipeline

Differential Quantitative Analysis

For continuous signal comparison across tissues:

1. Download fold-change-over-control bigWig files for each tissue
2. Use deepTools multiBigwigSummary to compute genome-wide signal matrix
3. PCA and hierarchical clustering to verify tissues separate by biology (not batch)
4. Use deepTools plotHeatmap at tissue-specific peak sets to visualize differences

Disease vs Normal Comparison

1. Match disease and normal samples on demographics (age, sex, life_stage)
2. Use encode_search_experiments with treatment or biosample filters
3. Be aware that disease samples may have altered cell-type composition
4. Differential peaks may reflect composition changes, not regulatory rewiring

Developmental Time Course

1. Use life_stage filter to find experiments at different developmental stages
2. Order by developmental time (embryonic, child, adult)
3. Track which elements gain or lose activity over development
4. Roadmap Epigenomics (2015) provides reference developmental trajectories

Step 9: Document and Track with Provenance

Record the entire comparison design and results:

# Track all experiments in the comparison
encode_track_experiment(accession="ENCSR_tissue1", notes="Cross-tissue H3K27ac comparison - pancreas")
encode_track_experiment(accession="ENCSR_tissue2", notes="Cross-tissue H3K27ac comparison - liver")

# Log any derived comparison files
encode_log_derived_file(
    file_path="/path/to/tissue_specific_peaks.bed",
    source_accessions=["ENCSR_tissue1", "ENCSR_tissue2"],
    description="Pancreas-specific H3K27ac peaks not found in liver",
    file_type="differential_peaks",
    tool_used="bedtools intersect v2.31.0",
    parameters="bedtools intersect -a pancreas.bed -b liver.bed -v"
)

# Link relevant publications
encode_link_reference(
    experiment_accession="ENCSR_tissue1",
    reference_type="doi",
    reference_id="10.1038/nature14248",
    description="Roadmap Epigenomics reference for cross-tissue comparison methodology"
)

Pitfalls and Common Mistakes

1. Confounding batch with biology: If all pancreas experiments come from Lab A and all liver experiments from Lab B, you cannot distinguish tissue differences from lab effects. Check lab metadata before interpreting any cross-tissue difference. Leek et al. (2010) showed that batch effects can dominate over 50% of total variation.

2. Mixing biosample types: Comparing K562 (cell line) H3K27ac with primary liver tissue H3K27ac conflates transformation-driven changes with tissue-specific regulation. Always compare within the same biosample type when possible.

3. Assembly mismatch: ALL files in a comparison must use the same genome assembly. GRCh38 and hg19 coordinates are NOT compatible. Use encode_compare_experiments to catch this.

4. Ignoring cell-type heterogeneity: Bulk tissue samples contain mixed cell populations. A "pancreas-specific" peak might actually be present in a minority cell type (e.g., delta cells). Single-cell data (scATAC-seq, scRNA-seq) can deconvolve this, but is not available for all tissues in ENCODE.

5. Depth-driven false differences: A deeply sequenced tissue will have more peaks called than a shallowly sequenced one, even if the underlying biology is identical. Always check sequencing depth and prefer normalized signal (fold change over control) for quantitative comparisons.

6. Incomplete panel comparison: Comparing 5 histone marks in tissue A but only 3 in tissue B produces a biased view. Document which marks are available in each tissue and restrict comparison to the intersection of available assays.

Walkthrough: Comparing Epigenomic Landscapes Between Normal and Cancer Tissues

Goal: Systematically compare ENCODE epigenomic data between normal tissue and cancer cell lines to identify disease-specific regulatory changes. Context: Comparing biosamples reveals which regulatory elements are gained or lost in disease states.

Step 1: Find experiments for both biosamples

encode_search_experiments(assay_title="Histone ChIP-seq", organ="liver", target="H3K27ac", organism="Homo sapiens")

Expected output:

{
  "total": 8,
  "results": [
    {"accession": "ENCSR100LIV", "biosample_summary": "liver", "target": "H3K27ac"},
    {"accession": "ENCSR200HEP", "biosample_summary": "HepG2", "target": "H3K27ac"}
  ]
}

Step 2: Compare the two experiments

encode_compare_experiments(accession_1="ENCSR100LIV", accession_2="ENCSR200HEP")

Expected output:

{
  "comparison": {
    "shared": {"assay": "Histone ChIP-seq", "target": "H3K27ac", "organism": "Homo sapiens"},
    "differences": {
      "biosample": ["liver", "HepG2"],
      "biosample_type": ["tissue", "cell line"]
    }
  }
}

Step 3: Download peak files for both

encode_download_files(accessions=["ENCFF100LIV", "ENCFF200HEP"], download_dir="/data/comparison")

Step 4: Identify differential peaks

bedtools intersect -v -a liver_peaks.bed -b hepg2_peaks.bed > liver_specific.bed
bedtools intersect -v -a hepg2_peaks.bed -b liver_peaks.bed > hepg2_specific.bed
bedtools intersect -a liver_peaks.bed -b hepg2_peaks.bed > shared_peaks.bed

Interpretation: HepG2-specific H3K27ac peaks mark cancer-gained enhancers. Liver-specific peaks mark enhancers lost in cancer.

Integration with downstream skills

• Differential peaks feed into peak-annotation for gene assignment
• Biosample-specific enhancers connect to disease-research
• Cell composition differences contextualized by cellxgene-context
• Expression differences validated via gtex-expression

Code Examples

1. Compare two experiments

encode_compare_experiments(accession_1="ENCSR100LIV", accession_2="ENCSR200HEP")

Expected output:

{
  "comparison": {
    "shared": {"assay": "Histone ChIP-seq", "target": "H3K27ac"},
    "differences": {"biosample": ["liver", "HepG2"]}
  }
}

2. Find matching experiments across biosamples

encode_get_facets(assay_title="Histone ChIP-seq", facet_field="biosample_ontology.term_name", target="H3K27ac", organism="Homo sapiens")

Expected output:

{
  "facets": {
    "biosample_ontology.term_name": {"K562": 8, "GM12878": 7, "HepG2": 5, "liver": 4, "brain": 3}
  }
}

3. Track compared experiments

encode_track_experiment(accession="ENCSR100LIV", notes="Liver H3K27ac - normal tissue control for HepG2 comparison")

Expected output:

{"status": "tracked", "accession": "ENCSR100LIV", "notes": "Liver H3K27ac - normal tissue control for HepG2 comparison"}

Integration

This skill produces...	Feed into...	Purpose
Differential peak sets	peak-annotation	Assign biosample-specific peaks to genes
Biosample-specific enhancers	disease-research	Identify disease-gained/lost regulatory elements
Shared regulatory elements	regulatory-elements	Define constitutive vs. tissue-specific cCREs
Cell composition context	cellxgene-context	Deconvolve tissue heterogeneity effects
Expression differences	gtex-expression	Validate regulatory changes with expression data
Comparison metadata	data-provenance	Document biosample comparison analysis
Differential regions	variant-annotation	Find variants in biosample-specific regulatory elements

Related Skills

• integrative-analysis: Combine multiple data types (ChIP + RNA-seq + ATAC) within or across biosamples
• epigenome-profiling: Build comprehensive epigenomic profiles for individual biosamples before comparing them
• quality-assessment: Evaluate data quality for each experiment BEFORE including it in a cross-tissue comparison
• histone-aggregation: Aggregate peaks across multiple experiments for the SAME tissue before cross-tissue comparison
• accessibility-aggregation: Aggregate ATAC-seq/DNase-seq peaks across experiments for the SAME tissue
• methylation-aggregation: Aggregate DNA methylation data across experiments for cross-tissue methylation comparison
• regulatory-elements: Use tissue-specific peak sets to discover enhancers, promoters, and other cis-regulatory elements
• single-cell-encode: Cell type-resolved data can deconvolve bulk cross-tissue comparisons
• multi-omics-integration: Combine multiple ENCODE assay types for deeper regulatory characterization
• data-provenance: Document all comparison parameters, tool versions, and results for reproducibility
• publication-trust: Verify literature claims backing analytical decisions

Presenting Results

• Present comparison as: metric | biosample_1 | biosample_2 | difference. Highlight significant differences. Include compatibility verdict (COMPATIBLE/WARNING/INCOMPATIBLE). Suggest: "Would you like to download the compatible experiments?"

For the request: "$ARGUMENTS"