henhaobio
Sign in

Genome biologyBioinformaticsPeer-reviewed journal article

Model-based Analysis of ChIP-Seq (MACS)

Yong ZhangiD, Tao LiuiD, Clifford A. MeyeriD, Jérôme EeckhouteiD, David S. JohnsoniD, B BernsteiniD, Chad NusbaumiD, R MyersiD, Myles A BrowniD, Wei LiiD, X. Shirley LiuiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

We present Model-based Analysis of ChIP-Seq data, MACS, which analyzes data generated by short read sequencers such as Solexa's Genome Analyzer. MACS empirically models the shift size of ChIP-Seq tags, and uses it to improve the spatial resolution of predicted binding sites. MACS also uses a dynamic Poisson distribution to effectively capture local biases in the genome, allowing for more robust predictions. MACS compares favorably to existing ChIP-Seq peak-finding algorithms, and is freely available.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text (17 Sept 2008)

MACS calls ChIP-Seq peaks by estimating tag shift from Watson/Crick strand offsets and testing enrichment against a locally adaptive Poisson background, improving specificity and binding-site resolution.

ChIP-Seq tags mark fragment ends rather than bound sites with an often unknown offset, and regional sequencing, mapping, chromatin, and copy-number biases were seldom modeled: of three recent studies with controls, only one used them systematically, and small peaks contain too few control tags to gauge background. MACS infers the offset from strand-specific tags and tests peaks against a regional Poisson rate, working with or without a matched control.

Key results

  1. ShownAligning strand-separated tags from high-confidence peaks learns the Watson-to-Crick distance d from the data rather than a user-supplied fragment size; d nearly matched the motif-anchored value for FoxA1 and exceeded motif-based values for NRSF and CTCF.

    FoxA1 (3.9 million uniquely mapped tags): d = 126 bp (63 bp shift) vs 122 bp by FKHR motif alignment, despite ~500 bp sonication and ~200 bp size selection; NRSF: 96 bp vs 70 bp motif estimate; CTCF: 76 bp vs 62 bp; model built from 1,000 randomly sampled high-confidence peaks. · Modeling the shift size of ChIP-Seq tags; Figure legends

  2. ShownA per-peak local Poisson rate (λlocal) sharply curbs false positives, including when no control library exists, and with d/2 tag shifting improves motif-centered specificity and summit resolution; FoxA1 peak centers are also enriched for estrogen receptor motifs.

    At 7,000 FoxA1 peaks, empirical FDR was 0.4% with control and λlocal; without a control it was 3.8% using ChIP-derived λlocal versus 41.2% using a global λBG; shifting plus λlocal raised FKHR-within-50-bp rates and cut summit-to-motif distance vs unshifted or uniform-λBG runs; central 300 bp of peaks showed 3.1-fold estrogen receptor element (half-site 2.7-fold) enrichment over the genome. · Model evaluation; Figure legends

  3. ShownOn FoxA1 and NRSF data, MACS gave fewer sample-swap false positives than ChIPSeq Peak Finder, FindPeaks, and QuEST, higher motif occurrence near peak centers, and finer localization in most cases; on CTCF it also outperformed the two callers that run without a control.

    FDR = control peaks/ChIP peaks with identical parameters after swapping samples (no CTCF FDR, lacking control); motif occurrence within 50 bp of peak center consistently higher; center-to-motif distance computed after excluding peaks with no motif within 150 bp; QuEST omitted for CTCF since it needs a control. Values appear only in figures. · Method comparisons; Figure legends

  4. ShownFoxA1 ChIP-Seq and ChIP-chip calls largely agree; many sequencing-only peaks reflect array tiling gaps or sub-threshold array signal, platform-specific peaks carry motifs at rates similar to shared peaks, and ChIP-Seq summits sit closer to the FKHR motif.

    65.4% of ChIP-Seq peaks (FDR 1%) were also ChIP-chip peaks (MAT, FDR <1%, fold-enrichment >2); of the 34.6% unique, 1,045 (13.3%) lay in untiled or partly tiled regions, leaving 21.4% truly platform-specific; 835 of these 1,684 were recovered at ChIP-chip FDR 5%, and their array signal exceeded background (Wilcoxon p <10^-320); ChIP-chip peaks were twice as wide on average. · Comparison of ChIP-Seq to ChIP-chip; Figure legends

  5. SuggestedTag subsampling indicates strongly enriched sites are saturated at current depth whereas weaker ones would gain from deeper sequencing; published histone data hint that ChIP-Seq under-samples compact chromatin, as repressive-mark efficiency drops with differentiation.

    Saturation tables from 90% to 20% of tags (FoxA1, 3.9 million tags; NRSF, 2.2 million): peaks >60-fold enriched were saturated, sites <40-fold could still be gained; in Mikkelsen et al. data, H3K4me3 ChIP-Seq efficiency stayed high during differentiation whereas H3K27me3 and H3K9me3 efficiency declined; short MACS d estimates suggested Solexa favors shorter fragments. · Discussion; Modeling the shift size of ChIP-Seq tags

Methods, data and reproducibility

MACS drops excess duplicate tags, estimates offset d from Watson/Crick tag modes in 1,000 sampled high-enrichment regions, shifts tags by d/2, scans 2d windows for Poisson enrichment over a genome-wide λBG, rescores candidates with a local λ from 1 kb, 5 kb, and 10 kb windows, calls summits, and estimates sample-swap FDR; it was evaluated on new MCF7 FoxA1 and published NRSF and CTCF ChIP-Seq.

  • FoxA1 ChIP-Seq and matched control generated for this study in MCF7 cells; Illumina/Solexa 1G, two lanes each, DNA pooled from three independent experiments · human · n = 3.9 million ChIP and 5.2 million control uniquely mapped tags
  • FoxA1 ChIP-chip with controls in MCF7 cells, previously published; peaks called with MAT (FDR <1%, fold-enrichment >2) · human
  • NRSF ChIP-Seq with control in Jurkat T cells, from Johnson et al. · human · n = 2.2 million ChIP and 2.8 million control tags
  • NRSF ChIP-chip in Jurkat T cells from Gene Expression Omnibus · human · GSM210637
  • CTCF ChIP-Seq in CD4+ T cells from Barski et al., lacking a control sample · human · n = 2.9 million ChIP tags
  • Histone mark ChIP-Seq (H3K4me3, H3K27me3, H3K9me3) from Mikkelsen et al. across pluripotent and differentiating cells, used to gauge ChIP-Seq efficiency
  • Random 200 bp regions of the human genome excluding assembly gaps, used as background for motif occurrence · human · n = 100,000 regions
Methods & tools
MACS (Python command-line peak caller), ChIPSeq Peak Finder, FindPeaks, QuEST, MAT (ChIP-chip peak caller), Illumina/Solexa 1G Genome Analyzer, ChIP-Seq and ChIP-chip assays, FKHR and estrogen receptor element motif analysis, Poisson and binomial tag-count tests, Wilcoxon test
Reagents & environment
Python implementation (Artistic License), BED or ELAND input; defaults --gsize 2.7G bp, --tsize 25, --bw 300, --pvalue 1e-5, --mfold 32. Wet lab: MCF7 cells; ~3 ng ChIP and control DNA, PCR preamplification, 150 to 400 bp size selection.
Assay conditions & compute
2 GHz CPU Linux machine with 2 GB RAM: 15 seconds to model fragment size and under 3 minutes for peak calling on 3.9 million ChIP and 5.2 million control tags; sequencing on Illumina/Solexa 1G Genome Analyzer, two lanes per library.
Reproducibility gaps
Software stated as freely available but no URL appears in the text; no accession printed for the new FoxA1 ChIP-Seq or prior FoxA1 ChIP-chip; NRSF and CTCF ChIP-Seq referenced only by citation; Figures S1-S6 and Tables S1-S2 are in an external file.
Limitations (6)
  • When ChIP and control libraries differ in depth, the deeper sample tends to yield more peaks despite total-count scaling, so sample-swap FDR may be overly optimistic if ChIP tags outnumber control tags. · stated by the authors
  • In two-condition differential binding mode, sample swapping cannot estimate FDR because peaks in either sample may be genuine, so each sample's quality must be checked against a true input control. · stated by the authors
  • Repressive marks in compact chromatin may be under-recovered because such regions sonicate poorly and their longer fragments are disfavored by size selection; the authors advise harsher sonication or selecting larger fragments. · stated by the authors
  • The CTCF dataset lacks a control, so neither its FDR nor its fold-enrichment saturation could be estimated, and QuEST, which requires a control, was excluded from the CTCF comparison. · stated by the authors
  • Calling accuracy is inferred from motif proximity and sample-swap FDR on three human ChIP-Seq datasets, without orthogonal binding validation (e.g., ChIP-qPCR) or a description of the settings used for competing callers. · noted in this summary
  • A single genome-wide d/2 shift presumes punctate bimodal binding (d exceeded motif-based values for NRSF, 96 vs 70 bp, and CTCF, 76 vs 62 bp), and the one-parameter Poisson model is justified only at the depths tested; broad histone domains were not benchmarked. · noted in this summary

Read this if you call peaks from ChIP-Seq data and want to understand how fragment-shift estimation, local Poisson background, and sample-swap FDR shape binding-site resolution and false-positive control.

Citations
18,398
FWCI
53.98
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISVardaxis '18Awdeh '21Suryatengga '21Berger '19Teng '17Stanton '15
66% semantic matchBioinformatics
Ioannis Vardaxis, Finn Drabløs, Morten Beck Rye+1 more·Biostatistics·Dec 19, 2018· 4 citations