14 Combining histone marks
Combining histone marks means reading the changes in two marks together, at each gene. Two marks can tell you more together than each mark alone. H3K27ac marks active promoters and enhancers, whereas H3K27me3 marks genes that Polycomb proteins keep silent (Chapter 1). When a gene switches on during development, it often loses H3K27me3 and gains H3K27ac. This chapter shows how to read the two marks gene by gene, with an example from the developing pancreas.
- why H3K27ac and H3K27me3 are read together, and how this differs from calling bivalent promoters or poised enhancers;
- the three steps of a combined analysis: testing each mark, assigning regions to genes, and grouping the genes;
- how EpiNexus combines the two marks within one run;
- how to read the gene groups of the pancreas example (GSE84324);
- how a change in the cell mix, or too little power, can mislead the groups.
14.1 The question
Between two conditions, which genes lost the silencing mark and gained the active mark? Which genes changed in only one mark, and which in neither?
14.2 How it works
Two marks on one position
H3K27ac and H3K27me3 are two chemical groups on the same amino acid of histone H3, lysine 27. One position cannot carry both at once. In fruit flies, Tie and colleagues found that “H3K27ac and H3K27me3 are mutually exclusive” (Tie et al. 2009). In mouse cells, Pasini and colleagues described a switch: “The methylation to acetylation switch correlates with the transcriptional activation of PcG target genes” (Pasini et al. 2010). PcG means Polycomb group, the proteins that place H3K27me3.
Therefore, when a gene turns on, H3K27me3 is expected to go down at its promoter and H3K27ac to go up. When a gene turns off, the reverse is expected.
A bivalent promoter carries H3K27me3 together with H3K4 methylation, a mark of gene starts. Bernstein and colleagues found such promoters at developmental genes in embryonic stem cells. They proposed that these marks keep genes silent “while keeping them poised for activation” (Bernstein et al. 2006). A poised enhancer carries H3K4me1 and H3K27me3 but no H3K27ac (Rada-Iglesias et al. 2011). Neither can be called from H3K27ac and H3K27me3 alone.
Step 1: test each mark on its own
Each mark is first compared between the conditions on its own, as in Chapter 9. H3K27ac gives peaks of a few kilobases, whereas H3K27me3 covers broad domains and needs a method for broad marks (Chapter 7). Studies that combine the marks call each one with its own test and its own cut-off: for example, a false discovery rate (FDR) < 0.05 and a fold change above 2 for each mark (Lau-Corona et al. 2020), or FDR < 0.01 and a fold change above 1.5 (Basheer et al. 2019).
Step 2: give each changed region to a gene
Two rules are common:
- Promoter. A region belongs to a gene when it overlaps a window around the gene’s start. A window of 3 kb on each side is ChIPseeker’s default (Yu et al. 2015). Lau-Corona and colleagues gave an H3K27me3 site to a gene “if they overlapped the gene body or 3 kb surrounding its TSS” (Lau-Corona et al. 2020). TSS means transcription start site.
- Nearby. A region outside promoters goes to the closest gene start within a set distance. Basheer and colleagues assigned peaks “if they were within the 100-kb window from the transcriptional start site” (Basheer et al. 2019); Rada-Iglesias and colleagues used “a maximum distance of 100 kb” (Rada-Iglesias et al. 2011). This rule is a guess based on distance. The closest gene is often not the gene that an enhancer acts on (Chapter 13).
Step 3: group the genes
For each gene, each mark now has one call: gained, lost, unchanged (tested but not significant), or no region at all. Combining the two calls gives groups of genes. Lau-Corona and colleagues sorted the genes that rose in their knockout into “8 groups” by the marks around them (Lau-Corona et al. 2020).
Call each mark with its own test, at the cut-offs you would use for that mark alone. Do not relax the cut-offs for the second mark to make the groups bigger. Report how regions were assigned to genes: the window, and whether distant regions count.
From version 1.2.5, EpiNexus combines H3K27ac and H3K27me3 within one run.
- Put all samples of both marks, with their inputs, in one run. The run tests each mark on its own (Chapter 9). It then pairs the H3K27ac and H3K27me3 comparisons that have the same two conditions.
- Genes are the annotation’s protein-coding genes, with one start site each.
- Promoter calls: a mark is gained or lost at a gene when at least one region within 3 kb of its start changed significantly that way, by that mark’s own cut-offs. It is mixed when regions changed both ways, and unchanged when regions were tested but none changed.
- Groups come from the two promoter calls, for example “H3K27ac gained, H3K27me3 lost”.
- Nearby: a region that overlaps no promoter goes to the closest gene start within 100 kb. Nearby changes are shown beside the groups; they do not decide them.
The Integration page shows the groups, a flow chart from H3K27ac calls to H3K27me3 calls, and a table of the changed genes. You can download the full gene table. The report and the methods text describe the same result. A run with only one of the two marks skips the step and says why.
14.3 The example: becoming an endocrine cell
In the embryonic pancreas, progenitor cells that will make hormone-producing (endocrine) cells switch on the transcription factor Ngn3 (gene Neurog3). Cells with high Ngn3 have committed to the endocrine fate. Yu and colleagues sorted cells with low and high Ngn3 from the dorsal pancreas of mouse embryos at day 13.5, and profiled both marks (GEO GSE84324) (Yu et al. 2018). They used about 15,000 cells per ChIP, far fewer than a standard ChIP uses.
| Mark | Ngn3-low | Ngn3-high | Reads |
|---|---|---|---|
| H3K27ac | GSM2231448, GSM2231449 | GSM2231450, GSM2231451 | 16.0–17.4 million |
| H3K27me3 | GSM2231452, GSM2231453 | GSM2231454, GSM2231455 | 17.3–23.3 million |
| Input | GSM2231464, GSM2231467 | GSM2231466, GSM2231465 | 14.4–28.8 million |
Together, the files are about 11 GB. The study used the mm10 genome, and we therefore used it too.
- Create a project with the genome Mouse (mm10) and import the twelve GSM numbers.
- Check the replicate numbers before you import. GEO titles these samples “#1” and “#2”. Since version 1.2.7, EpiNexus reads them as replicate numbers. Inputs are matched to ChIP samples by condition and replicate. Therefore, check that each input has the replicate number of its ChIP samples. (Version 1.2.5, which made the run in this chapter, did not read “#1” and “#2”. We set GSM2231466 to replicate 1 and GSM2231465 to replicate 2 by hand.)
- The import names the conditions “dorsal pancreas; Ngn3-low” and “dorsal pancreas; Ngn3-high”. Rename them Ngn3-low and Ngn3-high on the Samples page.
- Run ChIP-seq, Full Pipeline, with the defaults. The Parameters step lists one comparison per mark, both Ngn3-high against Ngn3-low (Figure 14.1). Keep both.
14.4 Reading the result
Each mark on its own
In the run (b101ab08, EpiNexus 1.2.5), H3K27ac was tested on 122,888 regions from one peak call on the pooled libraries (Chapter 9). In Ngn3-high cells, 4,043 regions gained H3K27ac and 4,372 lost it (Figure 14.2). H3K27me3 was tested with csaw on 45,264 regions. Only 240 of them changed: 228 lost H3K27me3, and 12 gained it (Figure 14.3).
The two marks behaved differently. H3K27ac changed in both directions, at thousands of regions, whereas H3K27me3 changed at a few hundred regions and almost always went down. These results suggest that, as cells commit to the endocrine fate, silencing is removed from some genes but rarely added.
The two marks together
EpiNexus found 16,357 genes with a region of either mark at their promoter (Table 14.2, Figure 14.4).
| Group | Genes |
|---|---|
| H3K27ac gained, H3K27me3 lost | 67 |
| H3K27ac lost, H3K27me3 gained | 0 |
| H3K27ac gained only | 811 |
| H3K27ac lost only | 686 |
| H3K27me3 lost only | 83 |
| H3K27me3 gained only | 2 |
| Both gained, or both lost | 0 |
| Mixed changes | 3 |
| No change | 14,705 |
The group of interest is the first: 67 genes where the silencing mark was removed and the active mark added. EpiNexus listed them by the strength of the change, and the first was Neurog3 itself (Figure 14.5). The group also included Pax6, Insm1, Isl1, Rfx6, Myt1 and Fev, genes of transcription factors known to drive endocrine cell development. Two more genes, Neurod1 and Arx, lost H3K27me3 at the promoter but gained H3K27ac nearby, not at the promoter. They therefore fell in the group “H3K27me3 lost only”.
No gene went the other way, and no gene gained or lost both marks. This result fits the switch described above: at a promoter, one mark replaces the other.
The nearby counts show what the promoter groups miss. Outside promoters, H3K27ac was gained near 1,456 genes and lost near 1,505. H3K27me3 was lost near 63 genes and gained near 5. Many changes in enhancers lie far from any promoter.
What the other results add
The run’s other analyses described the same change (Chapter 11, Chapter 12). Since version 1.3.0, Gene Ontology (GO) enrichment counts a long region that was split into pieces only once. A new run may therefore give slightly different GO terms.
- Regions that gained H3K27ac carried motifs of the NEUROD1 family (E-box factors; q = 1.1 × 10−190) and of PTF1A. Their top GO terms were about the fusion of synaptic vesicles with the membrane. Endocrine cells release hormones with much of the same machinery as nerve cells.
- Regions that lost H3K27me3 gave one significant GO term for a biological process: “pancreatic A cell differentiation”. A cells make glucagon.
- Regions that lost H3K27ac gave a different result; see the next pitfall.
The genes with the largest H3K27ac losses included Spi1 (PU.1; log2 fold change −4.0) and Ptprc (CD45), which are genes of blood cells. The lost regions carried motifs of blood-cell factors such as SPIB, and their top GO terms were about B cells and lymphocytes. The cells sorted for low Ngn3 therefore most likely included some blood cells. A sorted population is a mix of cells. A region can “lose” a mark because the cells that carried it are fewer, not because their chromatin changed. Check the strongest genes of each group: do they belong to the cells you meant to study?
A group named for one mark means that the other mark showed no significant change at that promoter. With two replicates per condition, a real change can be missed, and broad marks are harder to test than sharp ones. Here, H3K27me3 changed significantly at only 240 of 45,264 regions. In addition, one H3K27me3 library had 45 % duplicate reads and failed quality control in this run (Chapter 6). Since version 1.2.7, a high duplicate rate alone gives a warning, not a failure. The group “H3K27ac gained only” may therefore include genes whose H3K27me3 went down too little to be called.
Compare the marks only when both were tested between the same two conditions, ideally in the same run with the same samples. EpiNexus combines marks only within one run. If you analysed H3K27ac and H3K27me3 in separate runs, run them again together.
Key points
- H3K27ac and H3K27me3 are mutually exclusive marks on the same amino acid of histone H3. A gene that switches on is therefore expected to lose H3K27me3 and gain H3K27ac at its promoter.
- Bivalent promoters and poised enhancers cannot be called from H3K27ac and H3K27me3 alone.
- Test each mark on its own, at its own cut-offs; then assign the changed regions to genes, and group the genes by the two calls.
- EpiNexus combines the two marks only within one run: the regions within 3 kb of a gene’s start decide its group, and the regions nearby (within 100 kb) are shown beside it.
- In the pancreas example, 67 genes gained H3K27ac and lost H3K27me3 at their promoter, with Neurog3 first, and no gene went the other way.
- A change in the mix of sorted cells, or too little power for a broad mark, can put genes in the wrong group.
14.5 Further reading
- Tie et al. (2009) and Pasini et al. (2010): H3K27ac and H3K27me3 as two alternative marks of one position.
- Bernstein et al. (2006) and Rada-Iglesias et al. (2011): bivalent promoters and poised enhancers, which need H3K4 methylation as well.
- Lau-Corona et al. (2020) and Basheer et al. (2019): two studies that combined the marks gene by gene.
- Yu et al. (2018): the pancreas study used as the example here.