Artifact Regions in ChIP-seq and ATAC-seq: Why the Strongest Peak in Your Data Can Belong to No Protein at All

Zetobit Bioinformatics Insight Series

Epigenomics & Peak Calling

Artifact Regions in ChIP-seq and ATAC-seq: Why the Strongest Peak in Your Data Can Belong to No Protein at All

A handful of regions in every genome accumulate enormous sequencing signal in every experiment — whatever antibody you used, whatever cell type, whatever you were looking for. A peak caller cannot tell that signal from a real binding event, so it reports these artifacts as your most confident peaks. Removing them is not cleanup; it is a prerequisite for the analysis meaning anything.

By Kanna Nandakumar, PhD · Zetobit

ChIP-seq and ATAC-seq turn a biochemical enrichment into a genome-wide signal track: reads pile up where a protein was bound or where chromatin was open, and a peak caller marks the pileups. The logic is clean — more reads mean more of whatever you enriched for. The failure hiding in that logic is that some pileups have nothing to do with your enrichment. A small set of genomic regions produce towering read stacks in essentially any sequencing experiment, and those stacks are the tallest, most statistically confident features in the data.

These are the regions the ENCODE project curated into the blacklist. Their defining property is what makes them so treacherous: blacklist regions contain high, unstructured signal regardless of cell type or protein target, which is precisely what marks them as false positives.¹ A real peak is specific — it appears because this factor bound here in this condition. An artifact peak is the opposite: it appears everywhere, in every pulldown, because its cause is the genome and the sequencing, not the biology.

Where the false signal comes from

The artifacts are not random noise; they have specific, understandable origins in the structure of the genome assembly. Artifacts often occur near assembly gaps and in genomic regions with low-copy repeat elements, and they show high ratios of multi-mapped to unique reads.² Repetitive and low-complexity sequence attracts reads from all over the genome that cannot be placed uniquely; regions adjacent to assembly gaps collect reads that have nowhere correct to go. The result is a pileup that looks, to a peak caller, exactly like a strong enrichment — a tall stack of reads over a defined locus.

Two properties make blacklist regions especially insidious. First, they are reproducible: the regions that give rise to artifact signal are invariant for a given species across developmental stage and tissue, even though the signal strength there varies between experiments.² Because the same regions light up every time, an artifact peak replicates across your samples and even across studies — the usual evidence for a real signal. Second, they are assembly-dependent: where these signals arise is sensitive to the genome build, so new blacklists are generated for each successive assembly.² A blacklist made for hg19 does not fully protect an hg38 alignment, which ties this directly to the reference-build theme running through this series — the artifact map is a property of a specific assembly, not a universal constant.

How much signal are we talking about?

The scale is the part that surprises people. Blacklist regions cover only a small fraction of the mappable genome, yet they capture a wildly disproportionate share of the reads. In the ENCODE human ChIP-seq data, 582 million of 2.5 billion uniquely aligning reads — over a fifth — mapped to blacklisted regions, even though those regions are a small fraction of the chromosome.¹ The local enrichment is staggering: signal in these regions runs up to roughly 6400× background levels.³ No genuine transcription-factor peak is 6400-fold over background. When a feature that extreme sits in your data, a peak caller ranking by signal will place it at the very top of the list.

This is also why the fraction of reads in blacklist regions works as a quality metric. ENCODE uses it exactly that way — some low-quality experiments had up to 87% of their reads falling into blacklisted regions.¹ An experiment where the majority of the signal is artifact is not a slightly noisy experiment; it is one whose peak list, uncorrected, is mostly fiction with a few real peaks buried inside.

The input control is the tell. A real peak rises in the ChIP track but not in the matched input; an artifact region spikes in both, because its signal has nothing to do with the pulldown. Comparing against input — and masking known blacklist regions — is what separates specific enrichment from the genome’s built-in noise.

The two defenses, and why both matter

There are two complementary protections, and they catch overlapping but not identical problems.

The first is the input (or IgG) control: a matched sample that went through the same sonication, size selection, and sequencing but without the specific enrichment. Its signal is the background you expect from chromatin accessibility, copy number, and mappability alone — and it spikes at artifact regions just as the ChIP does. Modern peak callers use it as the null: MACS models the reads in each window as Poisson with a local background rate, calling a peak only when the ChIP signal exceeds what the local background predicts.⁴ Where the input is already high — an artifact region — the bar rises, and the false peak is suppressed. Without an input control, the caller compares your signal only to a global average, and every artifact clears it.

The second is blacklist masking: removing reads that overlap the curated artifact regions before peak calling. The blacklist is built from mappability annotation (via tools like UMap) plus the top-signal regions identified in input samples, merged within a set distance — and ENCODE recommends masking these regions before running peak callers like MACS2.² This catches artifacts that even an input control might not fully suppress, and it is cheap: one intersection step against a published BED file for your exact assembly.

Input control and blacklist are not redundant. The input adapts to your specific sample’s background but requires you to have sequenced one; the blacklist encodes known artifact regions for your assembly but is a fixed list that can miss sample-specific noise. The defensible pipeline uses both — and matches the blacklist to the exact genome build the reads were aligned to.

The assay-specific wrinkle

The ENCODE blacklist was built from ChIP-seq, and it is often applied to newer assays as if artifacts were universal. They are not entirely. ATAC-seq is frequently run without an input control at all, because the extra sequencing is costly — which removes one of the two defenses and makes blacklist masking more important, not less.⁴ CUT&RUN goes further: its different biochemistry produces a partly different set of false-positive regions, and a CUT&RUN-specific suspect list was needed because those regions persist even after ENCODE blacklist removal and even when peak calling against a negative control.⁵ The general lesson is that an artifact map is specific to an assay and an assembly; borrowing one from a different technique leaves gaps exactly where the new technique’s artifacts live.

What to check in any peak-calling pipeline

• Mask a blacklist matched to your exact assembly. Use the ENCODE (or assay-specific) list for the build your reads were aligned to — an hg19 list does not fully cover hg38.²
• Call peaks against an input/IgG control whenever possible. It sets a local background that suppresses artifact spikes; without it, a global background lets them through.⁴
• Report the fraction of reads in blacklist regions as QC. A high fraction is a red flag for a low-quality library, not a detail to skip.¹
• Be suspicious of your very strongest peaks. Enrichment far above any plausible biological level (hundreds- to thousands-fold) is the signature of an artifact, not a spectacular binding site.³
• Use an assay-appropriate suspect list. For CUT&RUN, ATAC, or CUT&Tag, the ChIP blacklist alone may leave technique-specific artifacts behind.⁵

Peak calling is where an epigenomics experiment becomes a list of sites someone will interpret as biology — motifs, target genes, regulatory circuits. That makes it the worst place to let the genome’s built-in noise masquerade as signal. Artifact regions are not subtle: they are the loudest thing in the data. They are just loud for the wrong reason, and the entire discipline of blacklisting and input controls exists because a peak caller, left alone, will always mistake the loudest signal for the most important one.

The Takeaway

A handful of genomic regions pile up ultra-high signal in every experiment regardless of what you pulled down — and a peak caller reports them as your most confident peaks.

Why it misleads: these artifacts come from low-mappability, repeat-rich, and assembly-gap regions, run up to ~6400× background, and can hold over 20% of all reads (up to 87% in bad libraries). They’re reproducible across samples and studies — the usual evidence for a real signal — so they replicate convincingly while meaning nothing.

What to do: mask a blacklist matched to your exact genome build, call peaks against an input/IgG control, report the fraction of reads in blacklist regions as QC, and distrust any peak enriched hundreds- to thousands-fold over background.

The hard limit: an artifact map is specific to an assay and an assembly. The ChIP blacklist doesn’t fully cover ATAC-seq or CUT&RUN, and an hg19 list doesn’t cover hg38 — match the list to both the technique and the build, or the gaps are exactly where the artifacts live.

References

1. Amemiya HM, Kundaje A, Boyle AP. The ENCODE Blacklist: identification of problematic regions of the genome. Scientific Reports. 2019;9:9354. nature.com/articles/s41598-019-45839-z
2. Klasfeld S, Wagner D. Greenscreen decreases type I errors and increases true peak detection in genomic datasets including ChIP-seq. bioRxiv. 2022. (On artifact origins, invariance, assembly dependence, and blacklist construction.) biorxiv.org/content/10.1101/2022.02.27.482177
3. Amemiya HM, Kundaje A, Boyle AP. The ENCODE Blacklist (signal up to ~6400× background at artifact regions). Scientific Reports. 2019;9:9354. pubmed.ncbi.nlm.nih.gov/31249361
4. Unsupervised contrastive peak caller for ATAC-seq (on MACS Poisson local-background modeling and ATAC input controls). bioRxiv / Genome Research. 2023. ncbi.nlm.nih.gov/pmc/articles/PMC9881890
5. Nordin A, et al. The CUT&RUN suspect list of problematic regions of the genome. Genome Biology / PMC. 2023. ncbi.nlm.nih.gov/pmc/articles/PMC10416431

© Zetobit LLC · Bioinformatics Insight Series zetobit.com