15 Your regions in public chromatin data
A regulatory atlas is a collection of public chromatin data from many tissues, mapped onto one shared set of regulatory elements. Comparing your regions with an atlas shows what kind of elements they are and in which tissues they are normally active. This chapter explains how the comparison works and how to read it. It also shows how EpiNexus uses an atlas to check that each sample looks like the tissue it came from.
- what candidate cis-regulatory elements and chromatin states are, and how an atlas holds them for many tissues;
- why the comparison depends on the regions you compare against, and how the width of your regions can mislead it;
- how to read a table of tissues and chromatin states, element classes and genome-wide states;
- how tissue match ranks the tissues that a sample resembles, and when it gives no confident match;
- what public data show about the lost and gained H3K27ac regions of the heart dataset.
15.1 The question
A differential analysis gives you a list of regions that changed. What kind of regions are they: promoters, enhancers or something else? In which tissues are they normally active? And does each of your samples look like the tissue it came from?
15.2 How it works
Chromatin states
A chromatin state is a combination of histone marks that often occur together, such as H3K4me3 and H3K27ac at active promoters. ChromHMM learns such states from the marks of each sample with a hidden Markov model, and labels every 200 bp of the genome with one state (Ernst and Kellis 2012). For the mouse, van der Velde and colleagues learned 18 states from ten assays in 66 epigenomes of twelve tissues, sampled “at daily intervals from embryonic day 11.5 to birth” (van der Velde et al. 2021). For human, EpiMap labelled 833 biosamples (Boix et al. 2021).
Fine states are hard to reproduce. Foroozandeh Shahraki and colleagues found that chromatin-state annotations “show a large degree of disagreement, even when run with the same method on replicated data sets”. They added: “Much of this disagreement can be attributed to inconsequential factors such as the granularity of label definitions and mismatch in the boundaries of annotated elements” (Foroozandeh Shahraki et al. 2024). EpiNexus therefore folds the states into nine broader groups: active TSS, flanking TSS, active enhancer, weak or primed enhancer, transcribed, bivalent or poised, Polycomb repressed, heterochromatin and quiescent.
A second kind of annotation describes the genome once, for all cell types. Vu and Ernst built such a “full-stack” annotation for the mouse “based on 901 datasets assaying 14 chromatin marks in 26 cell or tissue types”, with 100 states in 16 groups (Vu and Ernst 2023). It tells you what a region usually is, not what it is in one tissue.
Accessibility in each tissue
ENCODE also records, for each element and each biosample, whether the element is accessible there. For this, ENCODE compares the element’s DNase-seq signal in that biosample with the signal at all other candidate sites there, as a Z-score: “A high Z-score or a max-Z value is defined as >1.64 throughout, and low otherwise” (ENCODE Project Consortium et al. 2020). Elements with a low signal in a biosample are labelled “Low Accessibility” there (Moore et al. 2026). The mouse atlas of EpiNexus holds these calls for 298 biosamples that were profiled with DNase-seq: adult, postnatal and fetal tissues, blood and immune cells, embryonic stem cells and cell lines.
The background decides the question
To say that your regions are rich in active enhancers, you need something to compare them with. This is the background, or universe. The guide to LOLA, a widely used tool for this kind of test, defines it as “the set of regions you tested for possible inclusion in your user sets”. It adds: “Changing the universe isn’t right or wrong, it just changes the question you are asking” (Sheffield 2026).
For gained and lost regions, the background is all regions that the differential analysis tested. The test then asks what is special about the regions that changed, as for Gene Ontology (GO) terms and motifs (Chapter 11, Chapter 12). For the peaks of one sample, there are no tested regions. EpiNexus then replaces each peak by the elements it overlaps, and compares these elements with all elements of the index.
For each reference biosample and each state group, the test counts how many of your regions are in that state there, and compares this count with the background. A region counts as being in a state when any element it overlaps is in that state. A one-sided hypergeometric test (Fisher’s exact test) gives the p-value, and the odds ratio gives the size of the effect (Sheffield and Bock 2016). On a log2 scale, an odds ratio of 0 means “as often as the background”, 1 means twice the odds, and −1 half the odds. The p-values are corrected for the many biosamples and groups tested.
Wide regions overlap more elements
Your regions and the background can differ in ways that change the result. A long region overlaps more elements than a short one, and so has more chances to touch any state. Kanduri and colleagues give this example: “If for instance, the case regions are larger than the control regions, they are more likely to overlap the reference regions only due to their length” (Kanduri et al. 2019). EpiNexus therefore repeats every test against a matched background: regions of the tested set that overlap the same number of elements, lie as often near a promoter, and have a similar width. A result that is significant in both tests is called robust. This follows their advice “to assess the consistency of conclusions with different choices of null model” (Kanduri et al. 2019).
Finally, EpiNexus summarises each tissue group by the median log2 odds ratio of its biosamples, and by the share of them that are robust.
Genome-wide states
States are read at the middle of each element. Broad domains, such as those of Polycomb and heterochromatin, contain few elements and are therefore under-sampled. The genome-wide annotation covers every base. For it, EpiNexus reports a fold enrichment: the share of your regions’ bases in a state, divided by the share of the background’s bases in that state. This is the fold enrichment that ChromHMM computes (Ernst and Kellis 2026).
Tissue match
Which tissue does a sample resemble? Sheffield and colleagues showed that the pattern of DNase hypersensitive sites identifies the tissue: “Our models predicted tissue type with >80% accuracy in leave-one-out experiments” (Sheffield et al. 2013).
Most regulatory elements are active in many tissues and say little about any one of them. EpiNexus therefore uses only the elements that vary between tissues: those active in at least 3 reference biosamples and in at most 95 % of them. For each biosample, it tests whether the sample’s elements are over-represented among the elements active there. A tissue group scores the median log2 odds ratio of its best 3 biosamples. The best tissue is reported as the match when it leads the next by at least 0.6, its best biosample has q < 0.05, and the sample has at least 200 variable elements. These thresholds are EpiNexus’s own, calibrated on samples of known tissue.
Otherwise, EpiNexus reports no confident match. This is a real outcome, not a failure. In the same study, brain tumour samples of cell types that the model had not seen “were generally not strongly assigned to any lineage (average maximum probability 34%)” (Sheffield et al. 2013).
Not every mark identifies a tissue. In the Roadmap Epigenomics data, the states linked to H3K4me1, a mark of enhancers, were “the most tissue specific, with 90% of instances present in at most 5–10 epigenomes”. In contrast, “active promoters (TssA) and transcribed states (Tx, TxWk) were highly constitutive” (Roadmap Epigenomics Consortium et al. 2015). A repressive mark can group samples by other features: “H3K27me3 signal in Polycomb-repressed states grouped five immortalized cell lines together … despite their T-cell, lung, cervical, leukaemia and hepatocellular origins” (Roadmap Epigenomics Consortium et al. 2015). EpiNexus therefore tests accessibility (ATAC-seq, DNase-seq) and the marks of active elements and enhancers: H3K27ac, H3K9ac, H3K4me2 and H3K4me1. It shows the ranking of H3K4me3 and of transcription factors as indicative only, and it does not test repressive marks.
- Compare changed regions with the regions that were tested (Sheffield 2026; Kanduri et al. 2019).
- Report the size of each effect, such as the odds ratio, beside its q-value.
- Check that your regions and the background have similar widths and promoter shares, or use a matched background (Kanduri et al. 2019).
- Report which atlas release and which reference data you compared with, and cite the sources.
- Treat the result as context, not as proof of how the regions are regulated.
- When the comparison runs. After a run of ChIP-seq, CUT&Tag, CUT&RUN or ATAC-seq on hg38, mm10 or mm39, the run page shows the line Regulatory Atlas (after the run). Runs that finish after an atlas release are compared by themselves. For older runs, click Compare with the atlas. The comparison uses the finished run; nothing is analysed again.
- Releases. The atlas comes in releases, such as 2026.10b. A newer release is never applied automatically: the page offers Compare again with release …. The mm39 atlas is converted from the mm10 atlas.
- The page. Open Regulatory Atlas in the sidebar. Choose a region set: the gained and lost regions of each comparison (Differential regions), or the peaks of each sample (Sample peaks). When a release holds more than one reference collection, choose it under Reference. The tabs are Tissues × states, Element classes, Genome-wide states, Regions and Tissue match.
- The tissue check. On the Samples page, set each sample’s Reference tissue. The Tissue match tab then compares the match with it: Consistent, May not match, Unclear or Not checked. A sample that may not match is a reason to check for a swap or a mislabel. It is never a QC failure.
- Sources and output. The Data sources page lists every source with its version and licence. The report has a Regulatory Atlas section, and the export type Regulatory Atlas holds the tables. The GEO submission package never holds them, because they describe public data, not yours.
15.3 Reading the result: GSE155100
We compared run cafc042f of the heart dataset (Chapter 16) with release 2026.10b of the mouse atlas, on 18 October 2026, with EpiNexus 1.4.4. The run itself was analysed with EpiNexus 1.2.0; the comparison needed no new run. The release holds two reference collections: the chromatin states of 66 fetal and newborn tissue-stages, 7 of them heart, and ENCODE’s accessibility calls in 298 biosamples, 19 of them heart.
The lost regions are heart enhancers
The 2,267 lost pieces of Chapter 16 came from 2,211 regions. EpiNexus compared them with all 89,991 regions that the differential analysis tested, counted in the same way. Almost all lost regions (98.8 %) overlapped an element of the index.
Element classes. Distal enhancer-like elements were much more common in the lost regions than in the tested regions: 78.0 % against 52.1 % (q = 7 × 10−142; Table 15.2). Promoter-like elements were less common (10.4 % against 14.8 %). The H3K27ac that the knockout lost therefore lay mainly at enhancers, not at promoters.
| Class | Lost regions | Tested regions | Log2 odds ratio | q |
|---|---|---|---|---|
| Promoter-like (PLS) | 230 (10.4 %) | 13,283 (14.8 %) | −0.59 | 1.8 × 10−9 (fewer) |
| Proximal enhancer-like (pELS) | 154 (7.0 %) | 6,208 (6.9 %) | +0.02 | not significant |
| Distal enhancer-like (dELS) | 1,724 (78.0 %) | 46,907 (52.1 %) | +1.74 | 7.1 × 10−142 (more) |
| Accessible + H3K4me3 | 6 (0.3 %) | 694 (0.8 %) | −1.42 | 3.2 × 10−3 (fewer) |
| Accessible + CTCF | 4 (0.2 %) | 1,032 (1.1 %) | −2.54 | 3.2 × 10−7 (fewer) |
| Accessible + TF | 8 (0.4 %) | 375 (0.4 %) | −0.12 | not significant |
| Accessible only | 57 (2.6 %) | 5,574 (6.2 %) | −1.33 | 4.4 × 10−15 (fewer) |
| TF binding only | 1 (0.0 %) | 169 (0.2 %) | −1.50 | not significant |
| No element | 27 (1.2 %) | 15,749 (17.5 %) | −4.12 | 1.2 × 10−142 (fewer) |
Width. The lost regions were wider than the tested regions (median 1,894 bp against 687 bp) and overlapped more elements (3.76 against 2.23 on average). EpiNexus therefore also tested them against a matched background of 11,055 regions.
Chromatin states. In heart, the lost regions were active enhancers far more often than the tested regions: a median log2 odds ratio of 2.26, robust in all 7 heart tissue-stages (Figure 15.1). The strongest was the heart at birth, at 3.05. In the seven other tissue groups, the median for active enhancers was only 0.20 to 0.51. In these tissues, the same regions were instead more often transcribed (+1.17 to +1.36), Polycomb repressed (+0.56 to +1.11) or quiescent (+0.65 to +1.13). Active promoter states were rarer than in the background in every tissue. These results suggest that the lost regions are heart-specific enhancers: active in the heart, but not active enhancers in the other tissues.
Accessibility. ENCODE’s DNase data gave the same answer. In heart, the lost regions were accessible far more often than the tested regions: a median log2 odds ratio of 2.00, robust in 17 of 19 heart biosamples. No other tissue group came close: muscle was next, at 0.40 (robust in 4 of 16), and most groups were below 0 (Figure 15.2).
Single regions. The Regions tab labels each region. Of the 2,211 lost regions, 1,685 were most often active in heart. Two of them lay near Atp1a2, the gene followed in Chapter 16 and Chapter 8. The first was active in 58 of 66 tissue-stages, and its genome-wide state was “Promoters”. The second was active in only 9 of 66, most often in heart. It is therefore the more heart-specific of the two.
The gained regions sit in Polycomb chromatin
The 191 gained regions gave a different picture. No tissue stood out for active states, and no tissue group was robust in the accessibility data. In the genome-wide annotation, their bases were in Polycomb-repressed chromatin 3.02 times as often as the background, and in bivalent promoters 2.32 times as often. In the reference tissues, they were most often in Polycomb-repressed states, for example in lung (+1.73, robust). A possible reading is that, without Arid1a, H3K27ac appeared at some regions that Polycomb keeps silent in other tissues. With 191 regions, this is a lead to test, not a finding.
Each library resembled heart
All four libraries resembled heart, with a clear lead over kidney, the second tissue (Table 15.3). The best single biosamples were two postnatal hearts in ENCODE’s data, from mice 36 and 4 days old.
| Library | Variable elements | Heart | Kidney (next) | Lead |
|---|---|---|---|---|
| H3K27Ac_Arid1a_cKO_rep1 | 92,523 | 4.102 | 3.371 | 0.73 |
| H3K27Ac_Arid1a_cKO_rep2 | 104,838 | 4.027 | 3.274 | 0.75 |
| H3K27Ac_Control_rep1 | 98,400 | 4.117 | 3.222 | 0.90 |
| H3K27Ac_Control_rep2 | 112,620 | 4.138 | 3.248 | 0.89 |
Without the filter for variable elements, the picture is less clear. The elements under the peaks of the first control library were in active promoter states far more often than all elements, in every tissue: a median log2 odds ratio of 5.56 in heart, and 4.33 to 5.28 in the other groups. For active enhancers, heart led (2.74), but lung came close (2.16). Promoters are active in most tissues, and a library’s peaks hold many of them: 12.9 % of the elements under the peaks were promoter-like, against 3.0 % of all elements. The tissue signal lies in the elements that vary, which is why tissue match uses only them.
At first, no sample in this project had a Reference tissue, and the report said Not checked: no reference tissue given — set it on the Samples page. We then set the reference tissue of the four libraries to Heart on the Samples page. The Tissue match tab then read Consistent for all four (Figure 15.3). No new comparison was needed, because the check compares the stored match with the declared tissue each time the page opens.
Public atlases are built from public data. Your own dataset, or a closely related one, may be part of the reference, and a match with it is then expected and proves little. The exported Regulatory Atlas files list the result for every reference biosample, so you can see which ones a match rests on. Here, the best matches were ENCODE’s own heart samples, from mice 4 and 36 days old, not samples of this study.
The fetal states cover only eight tissue groups. A sample from another tissue, such as bone marrow, cannot match its own tissue there; at best it matches a relative. The DNase collection covers more tissues, but it holds no pancreas or skin. When the tissue is missing, expect no confident match, or the closest tissue that is present.
An atlas shows where your regions are normally active. It does not show which factor acts at them, or why they changed. Here it shows that the knockout lost H3K27ac at heart enhancers, not how Arid1a keeps these enhancers active.
Key points
- A regulatory atlas reads public chromatin data of many tissues at one shared set of elements, so that your regions can be compared with all tissues at once.
- The background decides the question: compare changed regions with all tested regions, and the peaks of a sample with all elements.
- Wide regions overlap more elements; a matched background shows whether a result holds when width and promoter share are equal.
- Tissue match uses only the elements that vary between tissues, and “no confident match” is a real outcome.
- In the heart dataset, the lost H3K27ac regions were heart-specific enhancers (median log2 odds ratio 2.26 for active enhancers in heart), and all four libraries resembled heart.
- An atlas gives context, not mechanism, and its reference may contain data like yours.
15.4 Further reading
- ENCODE Project Consortium et al. (2020) and Moore et al. (2026): the registry of candidate cis-regulatory elements.
- Ernst and Kellis (2012) and van der Velde et al. (2021): chromatin states, and those of the developing mouse.
- Vu and Ernst (2023): one chromatin-state annotation for all mouse cell types.
- Sheffield and Bock (2016) and Kanduri et al. (2019): how to test region sets against a background.
- Sheffield et al. (2013): tissue identity from regulatory activity.