GPTomics/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

| Tool | Method | Strength | Fails when | |------|--------|----------|------------| | ChromHMM (Ernst & Kellis 2012; v1.27 current) | Multivariate HMM on binarized 200 bp bins | Canonical; widely used; integrated with Roadmap Epigenomics 15-state model; mature toolchain | Binarization throws away signal quantitation; default 200 bp bins may be too coarse for sharp boundaries | | Segway (Hoffman 2012) | Dynamic Bayesian Network on continuous signal | Higher resolution; uses signal magnitudes not binarized | More complex setup; slower; less standardized output | | EpiSegMix (Schmitz, Aggarwal, Laufer, Walter, Salhab, Rahmann 2024 Bioinformatics 40:btae178) | HMM with flexible read-count distributions + duration modeling | Modern; handles both narrow and broad mark distributions in one model | Newer; smaller user base | | EpiLogos (Meuleman lab) | Multi-biosample visualization tool | Built on top of ChromHMM/Segway segmentations; compare ChromHMM states across 100s of biosamples | Visualization tool, not a segmentation method itself | | IDEAS (Zhang 2016) | Cell-type-aware joint inference | Across-cell-type segmentation respecting cell-type identity | Slower; complex parameter tuning | | EpiCSeg (Mammana 2015) | Negative binomial mixture | Read-count-based; doesn't need binarization | Less standardized output | | GenoSTAN | HMM with various emission distributions | Flexible | Less actively developed | | Roadmap 25-state model (Kundaje 2015) | ChromHMM 25-state precomputed model | Reference for cross-cell-type interpretation | Requires the Roadmap imputed 12-mark panel | | Full-stack ChromHMM (Vu Ernst 2022) | 100-state segmentation across 1032 datasets / 127 reference epigenomes | Comprehensive cross-tissue annotation | Computationally intensive to retrain |

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

| Roadmap 15-state assignments (canonical) | |-------------------------------------------| | 1_TssA — Active TSS (high H3K4me3, H3K27ac) | | 2_TssAFlnk — Flanking TSS (H3K4me3, H3K27ac, lower) | | 3_TxFlnk — Transcript flanking | | 4_Tx — Strong transcription (H3K36me3, H3K79me2 if available) | | 5_TxWk — Weak transcription | | 6_EnhG — Enhancer in gene body (H3K4me1, H3K27ac) | | 7_Enh — Generic enhancer (H3K4me1, H3K27ac) | | 8_ZNF/Rpts — Zinc-finger / repeats | | 9_Het — Heterochromatin (H3K9me3) | | 10_TssBiv — Bivalent TSS (H3K4me3 + H3K27me3) | | 11_BivFlnk — Bivalent flanking | | 12_EnhBiv — Bivalent enhancer (H3K4me1 + H3K27me3) | | 13_ReprPC — Polycomb-repressed (H3K27me3) | | 14_ReprPCWk — Weak Polycomb | | 15_Quies — Quiescent (no signal) |

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

| States | Use case | Mark panel size | |--------|----------|-----------------| | 8-10 | Initial exploration; small mark panel (3-4 marks) | 3-5 marks | | 15 | Roadmap Epigenomics canonical | 5 core (H3K4me3, H3K4me1, H3K36me3, H3K27me3, H3K9me3) | | 18 | Roadmap extended (adds H3K27ac -> fine enhancer subtypes) | 6 marks (core 5 + H3K27ac) | | 25 | Roadmap Epigenomics extended; cross-cell-type compatibility | 12 imputed marks | | 50+ | Full-stack model (Vu Ernst 2022) | Many marks across many cell types |

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.

Mark panel mismatch with model

Trigger: Applying Roadmap 25-state model to a sample with different mark panel.

Mechanism: Model was trained on specific marks; emission probabilities are mark-specific. Applying to different mark panel produces nonsensical state assignments.

Fix: Either train a new model on the available mark panel; OR ensure the exact same marks (and ordering) as used in the model.

IgG control instead of input

Trigger: Using IgG controls in BinarizeBam for histone marks.

Mechanism: ChromHMM's binarization compares mark signal to control; histone mark biology assumes input (sonicated chromatin) as background, not IgG.

Fix: Use sonicated input as control for histone mark ChIP. IgG is not appropriate for ChromHMM binarization of histone marks.

Cross-cell-type state mapping

Trigger: Training separate models per cell type and trying to compare state assignments.

Mechanism: State 5 in cell type A may not correspond to state 5 in cell type B if trained independently.

Fix: Train one model on concatenated data from all cell types (joint segmentation); or apply a single precomputed model (Roadmap 15-state, full-stack) to all samples for consistent state labels.

Reconciliation

| Pattern | Likely cause | Action | |---------|--------------|--------| | ChromHMM and Segway segments differ | Different bin sizes / binarization vs continuous | Both can be valid; inspect emission matrices; pick the tool matching the resolution needs | | State assignment varies wildly between replicates | Insufficient marks; over-binned | Increase mark panel; reduce state count | | Active TSS state overlaps polycomb state at promoters | Bivalent biology (Bernstein 2006) | Expected for ESC-like cells; not an error; consider bivalent-specific state in N=18 model | | Roadmap 25-state model annotates unknown cell type | Cross-cell-type generalization | Use cautiously; verify against tissue-specific tracks | | Heterochromatin (H3K9me3) state has too many bins | H3K9me3 covers large fraction of genome | Expected; heterochromatin is genome-wide |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | java.lang.OutOfMemoryError | Insufficient JVM heap | java -mx32G -jar ChromHMM.jar ... | | BinarizeBam very slow | Large BAMs without index | samtools index all BAMs first | | All states have similar emissions | Mark panel too small | Need at least 5 marks for canonical 15-state model | | Segments file empty for some chromosomes | Chromosome not in chromsizes file | Add or use -chrom flag to restrict | | State labels don't match Roadmap | Trained model independently | Use Roadmap precomputed model OR map states by emission similarity | | ChromHMM "no signal in marks" | All bins binarized to 0 | Check signal quality; verify control normalization |

References

• Ernst J & Kellis M 2012 Nat Methods 9:215 (ChromHMM v1)
• Ernst J & Kellis M 2017 Nat Protoc 12:2478 (ChromHMM protocol)
• Hoffman MM et al 2012 Nat Methods 9:473 (Segway)
• Schmitz JE, Aggarwal N, Laufer L, Walter J, Salhab A, Rahmann S 2024 Bioinformatics 40:btae178 (EpiSegMix)
• Meuleman W et al 2020 Nature 584:244 (EpiLogos / DHS index)
• Zhang Y & Hardison 2016 Nucleic Acids Res 44:6721 (IDEAS)
• Mammana A & Chung HR 2015 Genome Biol 16:151 (EpiCSeg)
• Roadmap Epigenomics Consortium 2015 Nature 518:317 (Roadmap 25-state model)
• Vu H & Ernst J 2022 Genome Biol 23:9 (full-stack ChromHMM)
• Kundaje A et al 2015 Nature 518:317 (Roadmap integrative analysis)

Related Skills

• chip-seq/peak-calling - Peak calling per mark before segmentation
• chip-seq/chipseq-qc - Replicate concordance per mark; QC before integration
• chip-seq/peak-annotation - cCRE classification (PLS/pELS/dELS) complementary to chromatin states
• chip-seq/spike-in-normalization - Spike-in normalize per-mark BAMs before binarization for cross-condition state comparison
• atac-seq/single-cell-atac - scATAC + ChIP integration via multimodal methods
• machine-learning/model-validation - Model selection (state count); cross-validation
• data-visualization/genome-tracks - Visualize state segmentations
• gene-regulatory-networks/coexpression-networks - Cross-reference chromatin states with co-expression modules