← All posts
by EpiNexus Team quality controlChIP-seqCUT&Tag

Is my ChIP-seq or CUT&Tag data good enough?

Before you trust a differential result, check each library on three things: is it deep enough, is it complex enough, and is the signal enriched? The first answer is often misread, because "enough reads" depends on what you profiled. ENCODE asks for 45 million usable fragments for a broad mark such as H3K27me3, but 3 million can be enough for a CUT&Tag library. And only a few results mean that a library has actually failed.

1. Depth: count usable fragments, not reads

Standards count usable fragments, not the reads that came off the sequencer. A usable fragment is one DNA molecule whose reads align with good mapping quality, are not duplicates, and don't lie in a blacklisted region. With paired-end reads, the two reads of a fragment count once. A library of 20 million reads with a 70 % mapping rate and 40 % duplicates therefore has far fewer than 10 million usable fragments.

How many you need depends on the assay and the target:

What you profiled Good Minimum Error Source
ChIP-seq: transcription factor, or a narrow histone mark (H3K27ac, H3K4me3, H3K9ac) 20 M 10 M under 5 M ENCODE
ChIP-seq: broad histone mark (H3K27me3, H3K36me3, H3K4me1) 45 M 35 M under 5 M ENCODE
ATAC-seq 25 M 20 M under 15 M ENCODE
CUT&Tag 3 M 1 M under 0.2 M mapped reads Kaya-Okur 2020; Henikoff lab tutorial; Murray 2026
CUT&RUN 3 M 1 M — Meers 2019; the minimum is an EpiNexus choice

All values are usable fragments, except the CUT&Tag error line, which a recent benchmark set in mapped reads. H3K9me3 is an exception in ChIP-seq: because it lies mostly in repeats, ENCODE counts all its mapped reads.

CUT&Tag and CUT&RUN need far fewer fragments than ChIP-seq because they have very little background. The developers of CUT&RUN write that they "sequence only 3-5 million fragments, because deeper sequencing provides little if any gain".

Too few fragments make you miss weak changes. They don't make the regions you do find wrong.

2. Complexity: are there enough different molecules?

A library made from too little material holds few distinct DNA molecules. Sequencing it more deeply only reads the same molecules again: the duplicate rate rises, and the number of usable fragments doesn't.

ENCODE measures complexity with three numbers:

  • NRF: distinct reads divided by all reads.
  • PBC1 and PBC2: how many positions carry exactly one read, compared with how many carry one or more (PBC1) or exactly two (PBC2).

For ChIP-seq, an NRF of 0.8 or more is good and below 0.5 is below ENCODE's minimum. For ATAC-seq the lines are 0.9 and 0.7.

CUT&Tag and CUT&RUN are different. The enzyme is tethered to the antibody, so it cuts at the same positions in many cells, and identical fragments can be real signal. EpiNexus keeps them in these assays, and shows their duplicate rate for information only.

3. Enrichment: do reads gather where they should?

FRiP, the fraction of reads in peaks, is the most common measure, and the one most often misused.

  • Transcription factors in ChIP-seq: ENCODE looks more closely at any experiment whose FRiP is below 1 % (Landt et al. 2012).
  • Histone marks in ChIP-seq: there is no published FRiP value. FRiP depends strongly on how peaks are called, so compare it only between libraries of the same target that were analysed together.
  • Histone marks in CUT&Tag and CUT&RUN: the nf-core/cutandrun pipeline calls a FRiP above 0.3 (30 %) reasonable.

For ChIP-seq of transcription factors and narrow marks, the strand cross-correlation scores NSC and RSC are a second check. ENCODE repeats replicates below 1.05 and 0.8.

What counts as a failure?

EpiNexus grades each value in one of four ways:

  • Good: at or better than the good value.
  • Marginal: between the good value and the minimum.
  • Below minimum: worse than the minimum.
  • Fail: worse than the fail value.

Only Fail fails a sample. A fail value exists only where ENCODE sets an error level, such as fewer than 5 million usable fragments for ChIP-seq. A library with no usable fragments, or with peaks but no reads in them, also fails. A library Below minimum is kept, with a warning.

Values that no publication sets, such as EpiNexus's lines for mapping rate, duplicates and peak counts, never fail a sample. This follows ENCODE's own practice: it looks more closely at libraries below its values, but doesn't discard them automatically. A marginal value is a reason to look, not a verdict.

Where to see it in EpiNexus

Since EpiNexus 1.2.7 the Quality Control page grades every library this way:

  • Dashboard tab: shows each metric, coloured by its grade.
  • Metrics tab: lists every value, with a Graded as column that says which lines apply to each library (for example "broad histone mark").
  • Interpretation tab: names each weak metric and suggests what to do.

You decide what happens to failing samples before a run starts. In the Advanced Parameters of the wizard's Parameters step, Samples that fail QC either keeps them and flags the results (the default) or excludes them from the statistics. Excluding leaves out only samples graded Fail.

What to do with a weak library

The metric that is weak usually points to the cause:

What is weak Likely cause What to do
Depth only, complexity good not sequenced deeply enough sequence the same library again
Complexity too little material or too many PCR cycles make a new library; more sequencing won't help
Enrichment weak antibody or pull-down repeat the experiment
Mapping rate contamination, wrong reference genome, adapters, a large spike-in check the species, genome, trimming and spike-in

With only two replicates per group, think twice before removing one: the group is then left with a single library, and you can no longer see how much replicates differ. Often it's better to keep a marginal library and report it.

The book covers each metric in depth, with a worked example: Quality control you can trust. Every threshold and its source is listed in Appendix B. EpiNexus is in an invite-only pilot: ask for access.

Sources