Written and maintained by CASRAI Editorial Board
Last updated
ChIP-seq maps where a specific protein sits on the genome. It answers the question accessibility assays cannot: not “is this region open” but “is this protein here”. That specificity comes entirely from an antibody, which is also the reason ChIP-seq experiments fail more often, and more quietly, than most sequencing assays.
The four steps
- Crosslinking. Formaldehyde reversibly and covalently fixes chromatin-binding proteins to the DNA they occupy, freezing the interaction in place.
- Fragmentation. Chromatin is broken into manageable pieces, either by sonication or by micrococcal nuclease digestion, then purified away from the rest of the cell.
- Immunoprecipitation. An antibody against the target protein pulls down the fragments it occupies.
- Quantification. Crosslinks are reversed, DNA is purified, and the recovered fragments are measured by sequencing (or qPCR at individual loci).
Two of these are optimisation problems you should solve before running the real experiment: sonication conditions, which are cell-type and volume dependent, and antibody specificity.
The antibody is the experiment
Everything downstream inherits the antibody’s selectivity. A high-quality, ChIP-validated antibody is what keeps the signal-to-noise ratio high enough to call peaks at all; a poor one produces a library that sequences perfectly well and means nothing.
This is why “validated for western blot” is not the relevant credential. A western blot presents a denatured, size-separated protein; ChIP presents a native protein inside a crosslinked complex. An antibody can be excellent at one and useless at the other. Confirm ChIP-grade specificity before committing a sequencing run, not after.
Controls, and why input alone is not enough
An input control — sheared chromatin that skipped the immunoprecipitation — is mandatory. It captures the biases that have nothing to do with your protein: uneven fragmentation, mappability differences, copy-number variation. Peaks are called against it, not against a flat expectation.
But input normalisation has a documented limit. Library preparation uncouples enrichment from the amount of starting chromatin, so comparing global levels between samples — say, a histone mark before and after a treatment that changes it genome-wide — is not something input normalisation can do reliably. Every library is sequenced to a similar depth regardless of how much material went in, which flattens exactly the global difference you were trying to measure.
Spike-in normalisation is the countermeasure. A constant, low amount of exogenous reference chromatin — commonly Drosophila — is added to every sample before immunoprecipitation. Because each sample then contains the same internal reference, signals can be scaled by equalising spike-in reads, absorbing technical variation in ChIP efficiency, PCR, library prep and sequencing. For tissue ChIP-seq, where input chromatin amounts genuinely differ between samples, input-adjusted spike-in normalisation goes further and accounts for that too.
The practical rule: if your question is where a protein binds, input control is enough. If your question is how much binding changed between conditions, you need a spike-in, and retrofitting one after sequencing is impossible.
ENCODE quality metrics, and what they mean
Three metrics do most of the work, and two of them have published thresholds.
- NSC (normalised strand coefficient) — the ratio of the fragment-length cross-correlation peak to background cross-correlation. ENCODE recommends NSC ≥ 1.05 for typical transcription factors in sharp-peak mode.
- RSC (relative strand correlation) — the ratio of the fragment-length peak to the read-length peak. ENCODE recommends RSC ≥ 0.8 under the same conditions.
- FRiP (fraction of reads in peaks) — the share of reads falling inside called peaks. Useful for comparing the same antibody across cell lines, or different antibodies against the same factor.
The cross-correlation metrics work without calling peaks first: they plot the correlation between read densities on the plus and minus strands as one is shifted along the genome. Real ChIP signal produces a peak at the fragment length, because reads pile up on either side of a bound site.
The FRiP caveat is important and routinely ignored. FRiP is sensitive to how peaks were called — the algorithm, its parameters, its thresholds. Comparing FRiP values from differently-processed datasets is meaningless. Any comparison must use peaks called uniformly by a single algorithm with a single parameter set, which in practice means reprocessing other people’s data rather than quoting their published number.
Sharp and broad targets behave differently
The thresholds above are quoted for sharp-mode transcription factors. Broad domains — many histone modifications, for instance — spread signal over large regions rather than concentrating it at discrete sites, which changes both the peak-calling approach and how the QC numbers read. A broad mark can produce an unimpressive-looking cross-correlation and still be a perfectly good dataset. Match the metric and the caller to the shape of the target rather than applying transcription-factor standards universally.
ChIP-seq or ATAC-seq?
They answer different questions and are frequently confused because both produce peak files. ATAC-seq reports which regions are physically accessible, needs no antibody, and works from small inputs. ChIP-seq reports which protein occupies a region, requires a validated antibody, and needs considerably more material.
Use ATAC-seq to find candidate regulatory regions; use ChIP-seq when you need to name the occupant. Neither substitutes for the other: an open region need not be bound by the factor you have in mind, and a bound site need not fall in a broadly accessible region.
Frequently asked questions
Why formaldehyde?
It crosslinks protein to DNA covalently but reversibly, so binding is preserved through fragmentation and immunoprecipitation and can then be undone before sequencing.
Sonication or MNase?
Both are used. Sonication shears chromatin mechanically and needs per-cell-type optimisation; micrococcal nuclease digests it enzymatically. The choice affects fragment-size distribution and therefore peak resolution, so keep it constant across samples you intend to compare.
Do I really need an input control?
Yes. Without it you cannot separate genuine enrichment from fragmentation bias, mappability and copy-number effects, all of which produce peak-shaped artefacts.
When do I need a spike-in?
Whenever you want to compare global signal levels between conditions. Library preparation decouples enrichment from input amount, so input normalisation alone cannot measure a genome-wide increase or decrease.
What NSC and RSC should I aim for?
ENCODE recommends NSC ≥ 1.05 and RSC ≥ 0.8 for typical transcription factors in sharp mode. Treat those as the reference point for sharp targets and interpret broad marks against caller-appropriate expectations instead.
Can I compare my FRiP to a published one?
Only if both were derived from peaks called by the same algorithm with the same parameters. Otherwise the comparison measures the peak callers, not the experiments.
References
- ChIP-seq guidelines and practices of the ENCODE and modENCODE consortia — Genome Research
- ENCODE 2012 quality metrics for integrative analysis — ENCODE
- Profiling of transcription factor binding events by ChIP-seq — PMC5544034
- Appropriate normalization is critical to improve reproducibility of tissue ChIP-seq — PMC12880172
- Recent advances in ChIP-seq analysis: from quality management to whole-genome annotation — PMC5444249
