Bio Chipseq Chromatin State Segmentation

Version Compatibility

Reference examples tested with: ChromHMM 1.27+, Segway 3.0+, EpiSegMix 1.0+, EpiLogos (Meuleman lab), IDEAS 1.20+, samtools 1.19+, bedtools 2.31+. ChromHMM requires Java 8+; runs as java -mx<MEMORY> -jar ChromHMM.jar <command>.

Chromatin State Segmentation

"Integrate multiple histone modification ChIP-seq tracks into chromatin states" -> Learn a small set of recurring combinatorial patterns of histone marks (active promoter, active enhancer, poised enhancer, polycomb-repressed, heterochromatic, transcribed, etc.) and segment the genome by which state each region belongs to. Output: per-state genomic intervals, state-by-mark emission matrix, and state-state transition matrix.

• CLI (canonical): ChromHMM BinarizeBam -> LearnModel -> OverlapEnrichment / NeighborhoodEnrichment
• CLI (continuous signal): Segway train -> posterior -> annotate
• CLI (flexible distributions): EpiSegMix (2024)
• Visualization across biosamples: EpiLogos (Meuleman lab)
• Cell-type-aware joint: IDEAS

Chromatin state segmentation requires a panel of histone marks; minimum 4-5 marks (e.g., H3K4me3, H3K27ac, H3K4me1, H3K36me3, H3K27me3) for meaningful states. With fewer marks, simpler peak-based annotation (chipseq/peak-annotation) is more appropriate.

Tool Taxonomy

ChromHMM Workflow

ChromHMM is the de facto standard. The workflow has 4 stages:

Step 1: Binarize ChIP-seq signal

# Build cellMarkFileTable: cell_type<TAB>mark<TAB>file<TAB>(optional control)
cat > cellMarkFileTable.txt << EOF
GM12878	H3K4me3	gm12878_h3k4me3.bam	gm12878_input.bam
GM12878	H3K27me3	gm12878_h3k27me3.bam	gm12878_input.bam
GM12878	H3K27ac	gm12878_h3k27ac.bam	gm12878_input.bam
GM12878	H3K4me1	gm12878_h3k4me1.bam	gm12878_input.bam
GM12878	H3K36me3	gm12878_h3k36me3.bam	gm12878_input.bam
EOF

# Binarize BAMs into 200 bp bins; emission = whether mark exceeds Poisson threshold
java -mx16G -jar ChromHMM.jar BinarizeBam \
    -b 200 \
    chromsizes_hg38.txt \
    bam_dir/ \
    cellMarkFileTable.txt \
    binarized_output/

Output: per-chromosome _binary.txt files, one row per 200 bp bin, columns = marks, values 0/1.

Step 2: Learn model

# Train HMM with N states; common choices: 15, 18, 25
# 15 states: Ernst & Kellis 2011 model; canonical
# 18 states: extends with additional regulatory states
# 25 states: Roadmap Epigenomics extended model
java -mx16G -jar ChromHMM.jar LearnModel \
    -p 8 \
    binarized_output/ \
    model_15state/ \
    15 \
    hg38

# Output: model_15state.txt (emission + transition matrices),
# emissions_15.png (visualization), transitions_15.png,
# per-chromosome _segments.bed (state assignments)
# AND automatically runs OverlapEnrichment + NeighborhoodEnrichment

Step 3: Interpret states from emission matrix

Step 4: Functional enrichment of states

# OverlapEnrichment: enrichment of each state for external feature sets.
# Options (e.g. -labels) MUST precede the three positional args.
java -mx16G -jar ChromHMM.jar OverlapEnrichment \
    -labels \
    model_15state/GM12878_15_segments.bed \
    /path/to/anchor_files/ \
    enrichment_output/GM12878

# NeighborhoodEnrichment: enrichment relative to anchor positions (e.g., TSS)
java -mx16G -jar ChromHMM.jar NeighborhoodEnrichment \
    -labels \
    model_15state/GM12878_15_segments.bed \
    /path/to/tss_anchors.txt \
    enrichment_output/GM12878_TSS

Anchor files: BED files of features (CGIs, repeats, conserved elements, etc.) for OverlapEnrichment; position files for NeighborhoodEnrichment.

Choosing State Count

Practical workflow: Train at N=15, 18, 25; compare emission matrices; choose the smallest N where biology is interpretable. Higher N risks over-segmentation (state splitting random variation).

Segway Workflow

# Segway reads only the Genomedata format, so first pack the signal tracks
# (one track per bigWig) plus the genome sequence into an archive.
genomedata-load \
    -s hg38.fa \
    -t h3k4me3=h3k4me3.bw \
    -t h3k27ac=h3k27ac.bw \
    -t h3k4me1=h3k4me1.bw \
    -t h3k36me3=h3k36me3.bw \
    -t h3k27me3=h3k27me3.bw \
    signal.genomedata

# Train Segway model; GENOMEDATA and TRAINDIR are positional
segway train \
    --num-labels=25 \
    --num-instances=3 \
    --resolution=100 \
    signal.genomedata traindir/

# Posterior probabilities + hard-call annotation (GENOMEDATA TRAINDIR OUTDIR, positional)
segway posterior signal.genomedata traindir/ posteriordir/
segway annotate signal.genomedata traindir/ identifydir/

# Output: identifydir/segway.bed.gz (state assignments)

Segway uses continuous signal (from the genomedata archive) vs ChromHMM's binarized bins. Trade-off: more information per region (continuous) but more complex training.

EpiLogos Visualization

EpiLogos doesn't perform segmentation; it visualizes existing ChromHMM/Segway segmentations across many biosamples (epilogos.org).

# Use precomputed ChromHMM segmentations across multiple cell types
# Web interface: https://epilogos.altius.org/
# Local: github.com/meuleman/epilogos

Useful for: cross-cell-type comparison; identifying tissue-specific regulatory states; cohort-level chromatin landscape summaries.

Full-Stack ChromHMM (Vu Ernst 2022)

The full-stack model trained on 1032 datasets / 127 reference epigenomes:

# Use precomputed model from Ernst lab
# github.com/ernstlab/full_stack_ChromHMM_annotations
# Annotate new sample by applying model to binarized data
java -mx16G -jar ChromHMM.jar MakeSegmentation \
    full_stack_model_100states.txt \
    binarized_sample/ \
    full_stack_output/

Useful for: applying a comprehensive cross-tissue annotation to a new sample; comparing to canonical Roadmap states.

Per-Tool Failure Modes

ChromHMM -- Bin size 200 bp too coarse for sharp boundaries

Trigger: Studying TF binding boundaries or sharp enhancer transitions at 200 bp resolution.

Mechanism: ChromHMM default 200 bp bins; biology may shift within a bin.

Fix: Reduce to -b 100 or -b 50 (smaller bin); increases memory and compute time but improves boundary resolution. Re-train model at finer resolution.

ChromHMM -- Binarization throws away signal quantitation

Trigger: Distinguishing low- from high-signal regions of the same state.

Mechanism: ChromHMM binarizes each 200 bp bin to 0/1 per mark; state assignment uses combinatorial pattern, not magnitude.

Fix: Use Segway (continuous signal) or EpiSegMix (flexible distributions) for magnitude-aware segmentation.

ChromHMM / Segway -- Wrong state count

Trigger: Training with N=50 states on a 4-mark panel; or N=10 on a 7-mark panel.

Mechanism: Excess states fragment biology; insufficient states force unrelated regions into the same state.

Symptom: Emission matrix shows redundant states (multiple states with same emission profile) at high N; or biologically distinct regions lumped together at low N.

Fix: Train at N=15, 18, 25; inspect emission matrix similarity; choose the smallest N where states are interpretable as distinct biology.