15  Case study: CUT&RUN and CUT&Tag

This chapter follows two experiments through EpiNexus, from public files to biology: a CUT&Tag study of a histone mark, and a CUT&RUN study of a transcription factor. For each, it shows how to read the results: quality control, peaks, a differential analysis with a spike-in, genes, motifs and Gene Ontology.

15.1 CUT&Tag: H3K27me3 in mouse spermatogonia

The question

Spermatogonia are the cells in the testis that produce sperm. The undifferentiated ones keep the stem-cell pool going. When they start to differentiate, they begin to make a receptor called KIT: undifferentiated spermatogonia lack KIT, and differentiating spermatogonia carry it (Hu et al. 2024).

H3K27me3 marks genes that are kept silent by the Polycomb system (Chapter 1). We ask: what happens to H3K27me3 when spermatogonia start to differentiate? Does it change in a few places, or across the whole genome?

The dataset

The data come from a study by Hu and colleagues (GEO GSE221942) (Hu et al. 2024). They sorted the two kinds of spermatogonia from the testes of adult mice by their surface proteins, including KIT. Each CUT&Tag reaction held 10,000 mouse cells, an antibody against H3K27me3, and 20,000 fly cells as a spike-in. The libraries were sequenced paired-end, 2 × 150 bp. There is no IgG control. We use the four H3K27me3 libraries (Table 15.1).

Table 15.1: The four CUT&Tag libraries used in this chapter. Read pairs from the European Nucleotide Archive.

Sample	GEO	Condition	Read pairs
WT_CutTag_Undiff_H3K27me3_rep1	GSM6910296	undifferentiated	9.64 million
WT_CutTag_Undiff_H3K27me3_rep2	GSM6910297	undifferentiated	10.65 million
WT_CutTag_Diff_H3K27me3_rep1	GSM6910298	differentiating	9.61 million
WT_CutTag_Diff_H3K27me3_rep2	GSM6910299	differentiating	10.40 million

Together the files are about 4.5 GB. The study used the mouse genome mm10, so we use mm10 too.

NoteIn EpiNexus 1.2

1. Create a project with the genome Mouse (mm10).
2. On Upload Data, choose From a public repository and enter the four GSM numbers. In the import table, check the conditions: the two groups differ only by cell type. The run in this chapter used the labels Undiff_WT and Diff_WT.
3. Start a run: CUT&Tag, Full Pipeline, genome mm10. On the Parameters step, check that the comparison reads Diff_WT vs Undiff_WT, so that the undifferentiated cells are the reference. Use the swap button if it does not.
4. Under Advanced Parameters, keep Use spike-in normalization on, and choose Fruit Fly (dm6/BDGP6) as the spike-in genome instead of E. coli. The fly genome must be installed; if it is not listed, ask the EpiNexus team.
5. This run also ticked Remove duplicate reads from target samples, as the data’s authors processed their reads (Hu et al. 2024). The default keeps them (Chapter 5).

The other settings are the defaults: SEACR in stringent mode, which, without IgG, keeps the top 1 % of signal blocks (Chapter 7). The run used in this chapter is b44d28b1, made with EpiNexus 1.2.3.

Quality control

Table 15.2: Quality control of the four libraries, from EpiNexus’s report (run b44d28b1). Reads count both ends of each pair and leave out the fly reads. Mapped: share of these reads that aligned to the mouse genome.

Sample	Reads	Mapped	Duplicates	FRiP	Peaks
undifferentiated, rep 1	17.6 M	79.1 %	22.3 %	43.1 %	9,154
undifferentiated, rep 2	19.5 M	74.8 %	22.2 %	43.6 %	9,361
differentiating, rep 1	16.9 M	80.4 %	19.8 %	43.1 %	10,576
differentiating, rep 2	18.4 M	80.7 %	19.7 %	44.1 %	10,951

The libraries look good (Table 15.2):

• Depth. Each library has 17–20 million mouse reads. The CUT&Tag authors write that “typically 3 million paired-end reads suffice for nucleosome modifications” (Kaya-Okur et al. 2019).
• FRiP is 43–44 %: almost half of the reads fall in peaks. This is typical of CUT&Tag, which has very little background.
• Duplicates are 20–22 %. This run removed them.
• Mapping. The two undifferentiated libraries are below EpiNexus’s target of 80 % (79.1 % and 74.8 %), so EpiNexus gives them a warning, not a failure. The 80 % target is EpiNexus’s own (Chapter 6). Their other metrics are as good as those of the differentiating libraries.

The replicates agree well. In the sample PCA (Figure 15.1), the first component holds 98.1 % of the variation and separates the two cell types.

Figure 15.1: Sample PCA of the four libraries, from the spike-in normalised counts in the tested regions. The first component (98.1 % of the variation) separates the cell types. Figure made by EpiNexus 1.2.3 (run b44d28b1).

Peaks

SEACR found 9,154 to 10,951 peaks per sample. They are broad: the median peak is 4.9–5.6 kb wide, and the peaks of one sample cover 60–66 Mb of the genome. H3K27me3 forms wide domains, so this is expected. Peaks found in both replicates give 8,346 regions in undifferentiated cells and 9,451 in differentiating cells.

About half of the peaks (49 %) lie within 3 kb of a gene’s start. A further 32 % lie far from genes. In every sample, the top Gene Ontology terms of the peaks are about development, such as “dorsal spinal cord development” and “dorsal/ventral pattern formation”. Polycomb keeps genes for other tissues silent, and these terms are what that looks like.

WarningPitfall: peak counts are not amounts

The differentiating samples have more peaks than the undifferentiated ones, but less H3K27me3 (see below). Without IgG, SEACR keeps the top 1 % of signal blocks in each sample. So the number of peaks shows how a sample’s signal is split into blocks, not how much of the mark there is. To compare amounts, use spike-in scaled tracks and the differential test.

Differential analysis: a global loss

EpiNexus merged the peaks that were found in at least two samples. That gave 10,713 regions, with a median width of 6.4 kb. For broad marks, EpiNexus normally tests windows with csaw (Chapter 9). csaw has its own normalisation, which would replace the spike-in. So, with a spike-in, EpiNexus tests the merged peaks with DESeq2 and spike-in size factors.

The spike-in shows that the differentiating cells have about a quarter less H3K27me3 per fly cell. Chapter 10 explains the numbers. As a result, most regions lose the mark (Figure 15.2):

• 5,897 regions lost H3K27me3 and 456 gained it. 84.5 % of all regions have less of the mark in differentiating cells.
• The losses are moderate. The median lost region drops by log₂ 0.86, a little less than half. 90 % of lost regions lie between log₂ −1.76 and −0.54.
• The gains are fewer but larger. The median gained region rises by log₂ 0.85, and a few rise more than eightfold.

Figure 15.2: EpiNexus’s differential result for differentiating against undifferentiated cells. Each point is a region. Blue: lost H3K27me3 (5,897); orange: gained (456); grey: not significant (4,360). Dashed lines: the thresholds (FDR 0.1, |log₂ fold change| 0.5). Figure made by EpiNexus 1.2.3 (run b44d28b1).

NoteIn EpiNexus 1.2

With two replicates per group, the Differential page adds two notes. DESeq2 cannot find outlier samples, because that needs three or more per group. And the variation of each region is mostly estimated from the trend across all regions. Treat the p-values as approximate; more replicates would help.

Genes: what changed where

Table 15.3: Example regions. Genes: the nearest genes of the peaks in each region, from EpiNexus’s annotation. log₂ change: unshrunk estimate.

Genes	Region (mm10)	Size	log2 change	FDR	Call
Kit	chr5:75,572,903–75,581,131	8.2 kb	−4.63	8 × 10−49	lost
Dll1	chr17:15,366,359–15,383,633	17.3 kb	−2.27	5 × 10−68	lost
Hoxa1	chr6:52,154,640–52,193,231	38.6 kb	−1.45	2 × 10−58	lost
Hoxa10	chr6:52,194,751–52,283,552	88.8 kb	+0.34	2 × 10−10	not called
Fendrr	chr8:121,070,813–121,096,009	25.2 kb	+3.77	1 × 10−101	gained
Foxl1	chr8:121,112,825–121,133,454	20.6 kb	+3.05	7 × 10−104	gained

Some regions tell a clear story (Table 15.3):

• Kit loses almost all of its H3K27me3, about 25-fold. KIT is the receptor that the differentiating cells make, and by which they were sorted. Losing a silencing mark fits a gene that is switched on. But H3K27me3 alone does not show expression; RNA-seq of the same cells would.
• The Hoxa gene cluster behaves in two parts. The region around Hoxa1 loses the mark. The next region, around Hoxa10, barely changes. Its FDR is tiny (2 × 10⁻¹⁰), but its change is only log₂ 0.34, an increase of 27 %. That is below EpiNexus’s threshold of 0.5, so it is not called. In a large region with many reads, even a small change can have a very small FDR. This is why EpiNexus also asks for a minimum change (Chapter 9).
• Two neighbouring regions on chromosome 8 gain strongly, 8- to 14-fold: one around the gene Foxl1, the other around Fendrr. These are the two most significant regions of the whole comparison. They are among the few places where H3K27me3 increases.

Gene Ontology

EpiNexus tested the lost and gained regions for Gene Ontology terms (Chapter 11).

The lost regions give no term at all. This is not because they lack a theme. EpiNexus tests lost and gained regions against all tested regions. Here the lost regions are 5,897 of the 10,713 tested regions, 55 %. A set that is more than half of its background can be at most 1.8 times enriched for any term (10,713 / 5,897). EpiNexus calls a term only at two-fold enrichment or more. So no term could pass. When a mark is lost almost everywhere, the question “what is special about the lost regions?” has little left to answer.

The gained regions give 28 terms. The top one, “activation of protein kinase C activity”, is 14.8-fold enriched. But it rests on 12 regions near only three genes (Lep, Sez6 and Sez6l). As in Chapter 16, check the genes behind a top term. Terms with many genes are more convincing. “Formation of primary germ layer” comes from 25 regions near 18 genes, such as T, Eomes and Snai1. “Heart morphogenesis” comes from 44 regions near 29 genes, such as Hand1 and Tbx5. These are genes of early embryos and of other tissues. Gaining H3K27me3 would silence them further. This is a lead to test, not a finding: the gained set is small.

Motifs

EpiNexus does not test broad domains for motifs (Chapter 12). Regions several kilobases wide are too large for a motif scan to point to one binding site.

What this case study shows

• A spike-in reveals a global change. The libraries have the same number of reads in peaks, but the spike-in shows that differentiating cells carry about a quarter less H3K27me3. Without it, the loss would be hidden (Chapter 10).
• Peak counts are not amounts. The samples with less H3K27me3 had more peaks.
• A small FDR is not a big change. Large regions with many reads can be significant for small changes. Look at the fold change too.
• GO needs a background that is larger than the set. When most regions change in one direction, a test against the tested regions has nothing to find.
• Read GO terms through their genes.

15.2 CUT&RUN: the transcription factor MafB in macrophages

The question

MafB is a transcription factor that macrophages make in large amounts (Vanneste et al. 2026). A transcription factor binds short DNA sites, often in enhancers and promoters, and changes how nearby genes are used. We ask: where does MafB bind in macrophages? And does the binding disappear in macrophages that lack MafB?

This dataset also shows two things that often come with CUT&RUN: one IgG control shared by all samples, and a sample whose peak calling fails.

The dataset

The data come from a study by Vanneste and colleagues (GEO GSE287346) (Vanneste et al. 2026). They grew macrophages from the bone marrow of two kinds of mice: control mice (Mafb^fl/fl) and mice whose macrophages lack MafB (Lyz2^Cre Mafb^fl/fl). They profiled MafB with CUT&RUN, two replicates per genotype, and made one IgG control from control cells. The enzyme of CUT&RUN, pAG-MNase, is made in bacteria and “brings along a fixed amount of Escherichia coli DNA that can be used as spike-in DNA for CUT&RUN” (Hu et al. 2024). We use the five libraries in Table 15.4.

Table 15.4: The five CUT&RUN libraries used in this chapter. Read pairs (about 2 × 150 bp) from the European Nucleotide Archive.

Sample	GEO	Condition	Read pairs
BMDM_IgG	GSM9267457	control (IgG)	7.81 million
BMDM_WT_MafB2_Rep1	GSM9267460	control	6.82 million
BMDM_WT_MafB2_Rep2	GSM9267461	control	7.58 million
BMDM_KO_MafB2_Rep1	GSM9267466	without MafB	6.99 million
BMDM_KO_MafB2_Rep2	GSM9267467	without MafB	8.97 million

Together the files are about 5.4 GB. The study used mm10, so we use it too.

NoteIn EpiNexus 1.2

1. Create a project with the genome Mouse (mm10), and import the five GSM numbers.
2. In the import table, check that the IgG sample has the target IgG, and that the conditions are the two genotypes. The run in this chapter used the GEO labels Mafbfl/fl (Mafb WT) and Lyz2Cre Mafbfl/fl (Mafb KO). Because the first label contains WT, EpiNexus uses it as the reference.
3. Start a run: CUT&RUN, Full Pipeline, genome mm10, with the default settings: SEACR in stringent mode against the IgG, the E. coli spike-in, and target duplicates kept.
4. The Review step notes which samples share the IgG. Since EpiNexus 1.2.1, a sample without an IgG of its own uses the run’s IgG, as nf-core/cutandrun does (nf-core community 2026).

The run used in this chapter is 71a4d9cc, made with EpiNexus 1.2.3.

Figure 15.3 shows the Review step for these five libraries.

Figure 15.3: The Review step of the run wizard for the five CUT&RUN libraries. Under Design notes: BMDM_IgG is also the control for the second control replicate and, from the other condition, for both samples without MafB; the super-enhancer step will be skipped; and a sample needs at least 1,000 E. coli spike-in fragments to be calibrated. EpiNexus 1.2 interface; not covered by the book’s licence.

Quality control

Table 15.5: Quality control of the five libraries, from EpiNexus’s report (run 71a4d9cc). Reads count both ends of each pair and leave out the E. coli reads. E. coli: spike-in fragments counted.

Sample	Reads	Mapped	Duplicates	FRiP	Peaks	E. coli	Grade
IgG	15.6 M	91.9 %	27.5 %	—	—	19,207	pass
control, rep 1	13.5 M	95.3 %	22.1 %	7.9 %	9,281	13,195	fail
control, rep 2	15.0 M	95.4 %	23.2 %	7.6 %	10,124	12,941	fail
without MafB, rep 1	13.7 M	93.5 %	27.5 %	0.15 %	721	57,010	fail
without MafB, rep 2	17.4 M	92.7 %	29.2 %	14.0 %	268,517	106,016	warning

Depth and mapping are good: 13–17 million reads, 92–95 % mapped. But EpiNexus grades three of the four MafB samples as failing (Table 15.5, Figure 15.4). Look at why:

• The control samples fail on FRiP: 7.6–7.9 %, below EpiNexus’s fail value of 10 %. That value is EpiNexus’s own, chosen with histone marks in mind (Table 6.1). A transcription factor binds short sites. Here about 10,000 peaks cover 14–15 Mb, half a percent of the genome. A FRiP of 8 % is good for that. For transcription-factor ChIP-seq, ENCODE’s line is 1 % (Landt et al. 2012).
• The first sample without MafB fails on FRiP (0.15 %) and on its number of peaks (721). But cells without MafB should give almost no MafB peaks. This “failure” is the expected result.
• The second sample without MafB gets only a warning, although its peak calling failed (see below).

Figure 15.4: The top of the Dashboard tab of the QC page for the CUT&RUN run (71a4d9cc). Each bar is one library, coloured by grade. FRiP: the two control samples are poor and the second sample without MafB marginal; the first sample without MafB has almost no bar. Peak count: that second sample has 268,517 peaks, graded good. EpiNexus 1.2 interface; not covered by the book’s licence.

WarningPitfall: QC grades do not know what a sample should look like

A grade compares each sample with fixed lines. It cannot know that a knockout sample should be empty, or that a transcription factor covers less of the genome than a histone mark. Read every grade against what you expect from the sample.

E. coli fragments. The samples without MafB have four to eight times more E. coli fragments than the controls. When the antibody finds no MafB, it pulls down little mouse DNA, so the fixed E. coli DNA makes up a larger share of the library. This is what spike-in normalisation uses (Chapter 10).

Replicates. In the sample PCA, the first component (98.1 % of the variation) separates the two genotypes. The two control replicates sit together; the two samples without MafB are further apart.

Peaks

SEACR found 9,281 and 10,124 peaks in the two control samples, with a median width of 1.4 kb. 6,344 peaks were found in both replicates. About half of the peaks (53 %) lie within 3 kb of a gene’s start; 26–27 % lie in introns and 15–16 % far from genes. The top Gene Ontology terms of the control peaks are about immune responses, such as “response to lipopolysaccharide”, a molecule of bacterial walls. This fits macrophages, which defend the body against bacteria.

In the samples without MafB, the two replicates disagree. The first has 721 peaks. The second has 268,517.

WarningPitfall: a peak call that failed

The second sample without MafB has 370 times more peaks than the first, and they cover 105 Mb, about 4 % of the genome. A sample without the protein has almost nothing to find, and for this sample SEACR’s threshold ended up far too low. The signs are easy to see: a peak count far above the other replicate, and peaks that cover a large part of the genome. Do not use such a sample’s peaks. Its reads can still be counted in shared regions, as the differential test below does.

Figure Figure 15.5 shows the average signal around the summits of the control peaks. The control samples rise sharply at the summits; the samples without MafB stay low. The IgG does something unexpected: it drops to almost zero at the summits.

Figure 15.5: Average signal around the summits of the control peaks, in spike-in scaled tracks. Green and vermillion: control replicates; blue and yellow: samples without MafB; dashed: IgG. Figure made by EpiNexus 1.2.3 (run 71a4d9cc).

The dip comes from how SEACR chooses peaks (Meers et al. 2019). With an IgG, its last step is, in the words of its source code (version 1.3), “eliminating control-enriched features”: it removes every target peak that overlaps a stretch where the IgG itself is high. The peaks that remain lie where the IgG is low, so the IgG’s average dips at their centres. The dip is a result of the selection, not a problem with the data.

Differential binding: binding disappears without MafB

For a transcription factor, EpiNexus tests windows of 501 bp around the peak summits, found in at least two samples (Chapter 9). That gave 6,555 windows. The result is clear (Figure 15.6):

• 6,308 windows lost binding, and none gained it. The median window has log₂ 4.5 less signal without MafB, about 22 times less.
• The raw reads in the windows are already about four times lower without MafB (48,000–57,000 against 198,000–201,000 per sample). The spike-in size factors (1.62 and 3.02 against 0.38 and 0.37) widen the gap further.
• The 247 windows that do not change are mostly places that both samples without MafB called, and the controls did not. They are not MafB sites.

Figure 15.6: MA plot of the comparison, without MafB against control, with spike-in size factors. Blue: lost binding (6,308); grey: not significant (247). No window gained binding. Figure made by EpiNexus 1.2.3 (run 71a4d9cc).

NoteIn EpiNexus 1.2

SEACR’s summit is the middle of the strongest stretch of a peak. It can move between replicates. EpiNexus 1.2 keeps a 501 bp window only when two samples’ windows overlap, so a site whose two summits are more than 501 bp apart is not tested. In this run, 628 of the 6,344 peaks found in both control replicates (10 %) were left out for this reason.

Genes

Table 15.6: MafB sites near genes that define macrophages. Gene and position from EpiNexus’s annotation of the control peaks; log₂ change: unshrunk estimate.

Gene	Where	Window (mm10)	log2 change	FDR
C1qa	promoter	chr4:136,898,813–136,899,314	−5.84	8 × 10−10
Csf1r	intron	chr18:61,108,659–61,109,160	−5.33	2 × 10−11
Mertk	promoter	chr2:128,698,170–128,698,671	−5.58	3 × 10−9
Fcgr1	7.7 kb upstream	chr3:96,300,937–96,301,438	−5.78	4 × 10−13
Cd163	intron	chr6:124,321,544–124,322,045	−4.07	3 × 10−6
Spi1	2.5 kb into the gene	chr2:91,084,272–91,084,773	−5.46	3 × 10−10
Zeb2	promoter	chr2:45,117,136–45,117,637	−5.55	1 × 10−7

MafB binds near many genes that make a macrophage what it is (Table 15.6): the complement gene C1qa, the receptor Csf1r that macrophages need to grow, Mertk and Fcgr1, which help macrophages take up dying cells and antibody-coated particles, and Spi1, the gene for PU.1, the main transcription factor of macrophages. All lose binding without MafB, 17- to 57-fold. Binding alone does not show that MafB controls these genes; RNA-seq of the same cells would.

Motifs

In the peaks. EpiNexus scanned the 5,000 strongest peaks, all from the control samples, against shuffled copies of the same sequences (Chapter 12). The top motifs belong to PU.1 and other ETS factors, then IKZF1 and KLF factors. MafB’s own motif is enriched too (2.1-fold), but ranks 29th. A CUT&RUN for one factor finds the places where it binds, and other factors often bind the same places. In macrophages, the main one is PU.1.

In the changed regions. EpiNexus compares the lost windows with windows that did not change, matched for GC content. Here that test is not meaningful: 6,308 windows were lost and only 247 did not change, and those 247 are not MafB sites. The control set is too small and too different to match. The test reports GC-rich motifs such as PATZ1 and EGR1, which say more about the control than about MafB.

WarningPitfall: tests that need an unchanged background

Motif and Gene Ontology tests of changed regions compare them with regions that did not change. When almost every region changes, as with a knockout of the factor itself, there is no proper background left. Here the lost windows are 96 % of all windows. The Gene Ontology test of the lost windows found no term: a set that is 96 % of its background can be at most 1.04 times enriched. Look instead at motifs and terms in the peaks themselves.

What this case study shows

• Binding disappears without the factor. 6,308 of 6,555 windows lose MafB binding, about 22-fold, and none gain it.
• QC grades need judgement. An empty knockout sample “fails”; a failed peak call only gets a warning; a transcription factor’s FRiP is lower than a histone mark’s.
• A failed peak call shows in its numbers: hundreds of times more peaks than its replicate.
• One IgG can serve all samples. SEACR keeps only peaks where the IgG is low, which is why the IgG dips at the peaks.
• Some tests need an unchanged background. When almost everything changes, motif and Gene Ontology tests of the changed regions have nothing to compare with.

15.3 Further reading

• Hu et al. (2024): the study that produced these data, with H2AK119ub1 CUT&Tag and PRC1 mutants.
• Kaya-Okur et al. (2019): CUT&Tag, how it works and how much sequencing it needs.
• Zheng et al. (2020): the CUT&Tag analysis tutorial.
• Vanneste et al. (2026): the study that produced the CUT&RUN data, with MafB in many kinds of macrophages.
• Skene and Henikoff (2017): CUT&RUN, how it works.
• Meers et al. (2019): SEACR, the peak caller used for both experiments.

Hu, Mengwen, Yu-Han Yeh, So Maezawa, Toshinori Nakagawa, Shosei Yoshida, and Satoshi H. Namekawa. 2024. ‘PRC1 Directs PRC2-H3K27me3 Deposition to Shield Adult Spermatogonial Stem Cells from Differentiation’. Nucleic Acids Research 52 (5): 2306–22. https://doi.org/10.1093/nar/gkad1203.

Kaya-Okur, Hatice S., Steven J. Wu, Christine A. Codomo, Erica S. Pledger, Terri D. Bryson, et al. 2019. ‘CUT&Tag for Efficient Epigenomic Profiling of Small Samples and Single Cells’. Nature Communications 10 (1): 1930. https://doi.org/10.1038/s41467-019-09982-5.

Landt, Stephen G., Georgi K. Marinov, Anshul Kundaje, Pouya Kheradpour, Florencia Pauli, et al. 2012. ‘ChIP-seq Guidelines and Practices of the ENCODE and modENCODE Consortia’. Genome Research 22 (9): 1813–31. https://doi.org/10.1101/gr.136184.111.

Meers, Michael P., Dan Tenenbaum, and Steven Henikoff. 2019. ‘Peak Calling by Sparse Enrichment Analysis for CUT&RUN Chromatin Profiling’. Epigenetics & Chromatin 12: 42. https://doi.org/10.1186/s13072-019-0287-4.

nf-core community. 2026. Nf-Core/Cutandrun: Analysis Pipeline for CUT&RUN and CUT&Tag Experiments. Https://nf-co.re/cutandrun.

Skene, Peter J., and Steven Henikoff. 2017. ‘An Efficient Targeted Nuclease Strategy for High-Resolution Mapping of DNA Binding Sites’. eLife 6: e21856. https://doi.org/10.7554/eLife.21856.

Vanneste, Domien, Wen Peng, Zhuangzhuang Liu, Malik Hamaïdia, Raphaël La Rocca, et al. 2026. ‘MafB Is a Conserved Transcriptional Regulator of Macrophage Development and Functional Identity Across Tissues and Species’. Immunity 59 (3): 559–576.e12. https://doi.org/10.1016/j.immuni.2026.01.012.

Zheng, Ye, Kami Ahmad, and Steven Henikoff. 2020. CUT&Tag Data Processing and Analysis Tutorial. Protocols.io. https://doi.org/10.17504/protocols.io.bjk2kkye.