GPTomics/bio-chipseq-motif-analysis

Version Compatibility

Reference examples tested with: HOMER 4.11+, MEME suite 5.5+ (STREME replaces DREME from 5.4+), monaLisa 1.10+, JASPAR 2024 CORE, HOCOMOCO v12, BioPython 1.83+, bedtools 2.31+.

DREME was removed from MEME suite 5.4+; use STREME instead. Some tutorials still reference DREME — verify the installed version via meme --version. JASPAR 2026 (released late 2025) integrates 1259 BPNet ChIP models in a Deep Learning collection; the CORE collection remains the standard PWM source.

Motif Analysis on ChIP-seq Peaks

"Find enriched DNA binding motifs in my ChIP-seq peaks" -> Discover de novo motif patterns and test for known TF motif enrichment in peak sequences, with appropriate background to control for compositional and positional biases.

• CLI (HOMER, fast): findMotifsGenome.pl peaks.bed hg38 outdir/ -size 200 -p 8
• CLI (MEME-ChIP, comprehensive): meme-chip -db JASPAR.meme peaks.fa
• R (regression-based, selective enrichment): monaLisa::calcBinnedMotifEnrR(seqs, bins, pwms)
• CLI (deep-learning-derived motifs): TF-MoDISco on BPNet attribution scores (see chip-deep-learning)

Motif discovery is sensitive to background choice and peak quality. Hyper-ChIPable artifacts at rRNA / housekeeping loci often produce false-positive motifs (GC-rich or A-T-rich biases of those regions). Filter peaks against blacklists and inspect peak distribution before running motif discovery.

Tool Taxonomy

| Tool | Discovery type | Background handling | Strength | Fails when | |------|----------------|---------------------|----------|------------| | HOMER findMotifsGenome.pl | De novo + known | GC-matched genomic regions (auto) | Fast (multi-core); integrated vertebrate/insect/plant DBs; one-command full report | Background can include unmasked repeats producing motif artifacts; -size given slow; auto background may include peaks themselves | | MEME-ChIP | De novo (STREME, MEME) + central enrichment (CentriMo) + DB comparison (TOMTOM) + scanning (FIMO) | Markov order-2 from input; shuffled (preserves dinucleotide) | Comprehensive single command; rigorous statistics; HTML report | Slower; sequences must be 100-500 bp; central enrichment requires summit-centered peaks | | STREME (MEME 5.4+) | De novo (replaced DREME) | Markov order-2 | Bailey 2021 benchmark: more accurate than DREME/HOMER/MEME/Peak-motifs; handles 3-30 bp; scales to 100k+ sequences | Memory-hungry for very long sequences (>1 kb) | | MEME (classical) | De novo (long, gapped) | Markov | Long motifs; gapped motifs | Slow (no parallel); replaced by STREME for short motifs | | DREME | De novo (short) | Shuffled | Historical; small fast | Removed from MEME 5.4+; use STREME | | monaLisa (Stadler lab) | Binned enrichment regression | Native (binned scoring) | Modern; regression-based; selectivity (TF-specific in differential peaks) | R-only; less integrated with browsers | | AME (MEME suite) | Known motif differential | Matched background set required | Designed for two-set comparison (e.g., peaks vs. control regions) | Requires user-provided background set | | CentriMo | Known motif central enrichment | Auto from input | Tests positional enrichment relative to peak center | Requires summit-centered peaks (200-500 bp) | | FIMO | Motif scanning | Markov model | Genome-wide scanning at user-set p-value | Many false positives at p ≤ 1e-4; tighten to 1e-5 for whole-genome | | HOMER scanMotifGenomeWide.pl | Motif scanning | None | Genome-wide scanning at fixed score threshold | Less calibrated than FIMO; HOMER's PWM format | | RSAT peak-motifs | De novo + known | k-mer comparison | Web-server; multi-tool ensemble | Web limits; less reproducible from CLI | | TF-MoDISco | DL attribution-based | Implicit in model | Motifs from BPNet/chromBPNet attribution scores; captures soft motif syntax | Requires trained DL model; see chip-deep-learning |

Background Selection — The Biggest Source of Error

Motif enrichment p-values are conditional on the background distribution. Wrong background produces wrong motifs.

| Background | What it preserves | Use case | Limitation | |------------|-------------------|----------|------------| | GC-matched genomic regions | Mononucleotide composition; chromatin context | TF motifs; avoid GC-bias artifacts | Doesn't preserve dinucleotide (CpG, TpA) | | Dinucleotide-shuffled | CpG and TpA frequencies | Short motifs; avoiding repeat-derived artifacts | Doesn't capture genomic position context | | Markov order-2 (trinucleotide) | Trinucleotide context | Compositional control; STREME/MEME default | Doesn't capture chromatin context | | Peak-flanking sequences (±500 bp upstream/downstream of peak) | Local genomic context | When peak GC differs from genome | May contain shared regulatory motifs if peaks cluster | | Repeat-masked input | Sequence with repeats replaced by N | Avoid TE-derived motif artifacts | Loses motifs in evolved-from-repeat regulatory elements | | Input control peaks | Open-chromatin / artifact regions | TF discrimination from generic chromatin | Hard to obtain; controversial | | Differential set (AME) | Treatment-condition-specific peaks vs ctrl peaks | Differential motif enrichment | Requires a control peak set |

Practical default: STREME / MEME-ChIP with Markov order-2 (built-in default); HOMER with -mask flag (mask repeats); always inspect for repeat-derived motifs (e.g., Alu-derived AluY consensus, LINE motifs).

Window Around Summit Matters

Motif enrichment improves dramatically when sequences are summit-centered:

| Window | When | |--------|------| | ±100 bp (200 bp total) | Sharp TFs (CTCF, p53); summit reliably reflects motif position | | ±150-250 bp (300-500 bp) | Most TFs and sharp histones; balance of motif coverage and noise | | Full peak width (-size given) | Variable-width broad marks; computationally expensive | | Whole gene body (>1 kb) | Almost always wrong; dilutes motif signal |

For ChIP-seq broad histone marks (H3K27me3, H3K9me3) motif analysis is generally NOT informative — these marks reflect Polycomb / heterochromatin domains without sequence-specific binding. Motif analysis applies to TFs and to histone marks deposited by sequence-specific cofactors (H3K27ac partial, since BRD4 reads acetyl).

HOMER Workflow

# De novo + known motif discovery, repeat-masked, GC-matched background
findMotifsGenome.pl peaks.narrowPeak hg38 homer_out/ \
    -size 200 \
    -mask \
    -p 8

# With user-supplied background (e.g., control peaks or random genomic)
findMotifsGenome.pl peaks.narrowPeak hg38 homer_out/ \
    -size 200 -mask -p 8 \
    -bg background_peaks.bed

# Known motifs only (skip de novo; faster)
findMotifsGenome.pl peaks.narrowPeak hg38 homer_known_only/ \
    -size 200 -mask -nomotif

# Differential motif analysis: peaks gained in condition A vs condition B
findMotifsGenome.pl gained_in_A.bed hg38 differential_motifs/ \
    -size 200 -mask -bg gained_in_B.bed

HOMER output files:

• homerResults.html — de novo motifs ranked by significance
• knownResults.html — known motif enrichment
• homerMotifs.all.motifs — all de novo motifs (PWM format)
• knownResults.txt — tab-separated known motif stats

MEME-ChIP Workflow

# Center peaks to ±100 bp around summit (column 10 in narrowPeak)
awk 'BEGIN{OFS="\t"} {summit = $2 + $10; print $1, summit - 100, summit + 100, $4, $5, $6}' \
    peaks.narrowPeak > peaks_centered.bed
bedtools getfasta -fi hg38.fa -bed peaks_centered.bed -fo peaks_centered.fa

# Full MEME-ChIP analysis
meme-chip \
    -oc meme_chip_out/ \
    -db JASPAR2024_CORE_vertebrates_non-redundant_pfms_meme.txt \
    -meme-nmotifs 5 \
    -streme-nmotifs 10 \
    -minw 6 -maxw 20 \
    peaks_centered.fa

# MEME-ChIP runs: STREME (replaces DREME) + MEME + CentriMo + TOMTOM + FIMO
# CentriMo tests known motifs for central enrichment — strongest signal of
# direct binding vs. tethered/indirect binding

monaLisa Workflow (R, Regression-Based)

Goal: Test which TF motifs are enriched in specific bins of peaks (e.g., bins of differential log2FC, or bins of accessibility).

Approach: monaLisa builds a per-motif regression of peak signal on motif occurrence, controlling for GC content. Selective for TFs that discriminate between bins.

library(monaLisa)
library(JASPAR2024)
library(Biostrings)

# Load peaks and split into bins (e.g., quintiles of log2FC)
peaks <- rtracklayer::import('peaks.bed')
peaks$log2FC <- ...  # from differential analysis
bins <- bin(peaks$log2FC, binmode = 'equalN', nElement = 200)

# Get sequences around peak centers
seqs <- getSeq(BSgenome.Hsapiens.UCSC.hg38, resize(peaks, width = 500, fix = 'center'))

# Load JASPAR PWMs
pwms <- getMatrixSet(JASPAR2024, list(species = 9606, collection = 'CORE'))

# Compute binned motif enrichment with GC control
res <- calcBinnedMotifEnrR(seqs = seqs, bins = bins, pwmL = pwms, BPPARAM = MulticoreParam(8))

# Plot heatmap of motif enrichment vs bins
plotMotifHeatmaps(x = res, which.plots = c('log2enr', 'negLog10P'),
                   width = 1.8, maxEnr = 2, maxSig = 10)

Per-Tool Failure Modes

HOMER -- Background includes peaks themselves

Trigger: Running findMotifsGenome.pl without -bg on a large peak set covering >5% of genome.

Mechanism: HOMER auto-samples GC-matched genomic regions for background, which can overlap the peak set itself.

Symptom: Even known TF motif p-values are weak (>1e-3); de novo motifs less enriched than expected.

Fix: Supply explicit -bg background (e.g., random genomic intervals matching peak count and width) OR use MEME-ChIP with internal shuffled background.

HOMER / MEME -- Repeat-derived false-positive motifs

Trigger: Running on unmasked peaks; peaks cover transposable elements (Alu, LINE, LTR).

Mechanism: TEs contain over-represented k-mers that motif algorithms mistake for biology.

Symptom: Top de novo motif matches AluY consensus (~280 bp), LINE/L1, or LTR families.

Fix: Use -mask (HOMER) or pre-mask peak sequences with RepeatMasker; verify TOMTOM matches against legitimate TF databases.

STREME -- Memory failure on long sequences

Trigger: Running STREME on full-peak sequences (>1 kb each) with > 50k peaks.

Mechanism: STREME holds suffix structures in memory; long sequences explode RAM.

Fix: Resize peaks to ±100-250 bp summit-centered before STREME; or downsample peak count.

CentriMo -- No central enrichment due to wrong centering

Trigger: Using full peak coordinates (BED start, end) without recentering on summit.

Mechanism: Peak start coordinate is the left edge, not the summit; motif may be enriched near the summit but appears unenriched relative to the start.

Symptom: Known TF motifs show no central enrichment despite being clearly enriched overall.

Fix: Recenter sequences on summit: summit = start + summit_offset (narrowPeak column 10) before extracting FASTA.

FIMO -- Massive false positives at default p ≤ 1e-4

Trigger: Genome-wide scanning at FIMO default p-value threshold.

Mechanism: At p ≤ 1e-4, expect ~3M random matches in a 3 Gb genome; most are false positives.

Symptom: FIMO output has millions of motif "hits"; can't distinguish real binding.

Fix: Tighten to --thresh 1e-5 or stricter for genome-wide scans; or restrict scan to peaks: fimo --bgfile motif_bg motif.meme peaks.fa.

monaLisa -- GC bins not respected

Trigger: Running calcBinnedMotifEnrR without GC binning when peaks have systematic GC differences (e.g., promoters vs distal enhancers).

Mechanism: GC-rich motifs are over-enriched in GC-rich bins simply by chance.

Symptom: Top motifs are CpG-rich families (e.g., E2F, NRF1) regardless of biology.

Fix: Use background = 'genome' with GC-matched genomic background; or background = 'otherBins' with stratified GC.

Motif analysis on hyper-ChIPable regions

Trigger: Running motifs on peaks dominated by hyper-ChIPable artifacts (rRNA, tRNA, housekeeping).

Mechanism: These regions have systematic compositional biases (GC-rich, A-T-rich, repeat-derived); motifs from artifact peaks reflect compositional biology, not TF binding.

Symptom: Top de novo motif matches no known TF; high-GC or low-complexity consensus.

Fix: Apply blacklist + custom hyper-ChIPable filter (top-1% input signal) before motif analysis. See chipseq-qc.

Reconciliation: When Motif Methods Disagree

| Pattern | Likely cause | Action | |---------|--------------|--------| | HOMER finds motif X; MEME-ChIP misses | Different background; HOMER may have permissive background | Run MEME-ChIP with explicit GC-matched background; check | | MEME finds long gapped motif; STREME doesn't | MEME captures variable-length structure; STREME limited to 30 bp | Both are correct; report MEME for long motifs | | Top de novo motif doesn't match TOMTOM databases | Novel motif OR repeat artifact OR compositional artifact | Inspect peaks for repeats; check input control; could be genuine novel TF | | Known motif enriched but no de novo recovery | Insufficient enrichment for de novo; or motif is degenerate | Trust known motif enrichment; de novo needs strong signal | | Differential motif gained in treatment but TF expression unchanged | TF post-translational regulation (binding mode change without expression change) | Check ChIP signal at known TF target genes; not a contradiction |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | HOMER "configureHomer.pl genome not installed" | Genome not configured | perl configureHomer.pl -install hg38 (one-time) | | MEME "sequence too short" | Peaks < motif min width | Resize peaks to ≥ 200 bp | | MEME-ChIP "out of memory" | Too many long sequences | Resize peaks to ±100-250 bp; downsample | | No enriched motifs | Peak quality / hyper-ChIPable / wrong background | Check FRiP, filter blacklist, supply explicit background | | Top motif is GC-rich consensus | GC-bias in peaks not matched by background | GC-matched background (HOMER -bg or MEME shuffled with order-2) | | FIMO produces millions of hits | p-value threshold too loose | --thresh 1e-5 for whole-genome; restrict to peaks for finer p | | TOMTOM matches always say "no match" | Motif database species mismatch | Use vertebrates / insects / plants DB matching organism |

References

• Heinz S et al 2010 Mol Cell 38:576 (HOMER)
• Bailey TL & Elkan C 1994 Proc ISMB (MEME)
• Bailey TL 2021 Bioinformatics 37:2834 (STREME)
• Machanick P & Bailey TL 2011 Bioinformatics 27:1696 (MEME-ChIP)
• Bailey TL et al 2015 Nucleic Acids Res 43:W39 (MEME suite update)
• Grant CE et al 2011 Bioinformatics 27:1017 (FIMO)
• Bailey TL & Machanick P 2012 Nucleic Acids Res 40:e128 (CentriMo)
• McLeay RC & Bailey TL 2010 BMC Bioinformatics 11:165 (AME)
• Machlab D et al 2022 Nucleic Acids Res 50:e49 (monaLisa)
• Castro-Mondragon JA et al 2022 Nucleic Acids Res 50:D165 (JASPAR 2022; CORE collection)
• Avsec Ž et al 2021 Nat Genet 53:354 (BPNet; soft motif syntax)
• Shrikumar A et al 2020 bioRxiv (TF-MoDISco)

Related Skills

• chip-seq/peak-calling - Upstream peak calling; recenter on summit for motif input
• chip-seq/chipseq-qc - Filter hyper-ChIPable artifacts before motif discovery
• chip-seq/chip-deep-learning - BPNet/chromBPNet for sequence-attribution motif discovery (TF-MoDISco)
• chip-seq/peak-annotation - Annotate peaks before motif discovery to filter promoters vs enhancers
• atac-seq/motif-deviation - chromVAR per-cell motif activity (ATAC-specific)
• atac-seq/footprinting - TOBIAS footprint analysis (ATAC; complementary to motif enrichment)
• sequence-manipulation/motif-search - General sequence motif scanning
• genome-intervals/proximity-operations - bedtools getfasta to extract peak sequences