CpG Island Criteria Explained, With Worked Examples

Why CpG is rare, and where it is not

In vertebrate genomes most CpG cytosines are methylated, and 5-methylcytosine deaminates spontaneously to thymine. Over evolutionary time that has eroded CpG into TpG (and CpA on the opposite strand), so across the genome CpG turns up at only about a fifth to a quarter of the frequency its C and G content predicts.

CpG islands are the exception: stretches, often over promoters and first exons, that are kept free of methylation in the germline and so never lost their CpGs. That is why they are found by counting — an island is simply where the dinucleotide has survived.

The observed/expected CpG ratio

Observed/expected CpG = number of CpG ÷ ((number of C × number of G) ÷ length). The denominator is how many CpGs you would expect if C and G were scattered independently along the sequence; a ratio of 1 means no depletion at all. It is the formula Gardiner-Garden and Frommer used in 1987, and the one UCSC's CpG-island track documents.

Two things to get right. Count the dinucleotide CG read 5'→3' along one strand — CpG is its own reverse complement, so the count is identical on both strands and must not be doubled. And do not confuse it with GC content, which counts C and G bases individually and says nothing about whether they sit next to each other.

The two published definitions

• Gardiner-Garden and Frommer (Journal of Molecular Biology, 1987): longer than 200 bp, GC content above 50%, and observed/expected CpG above 0.6. This is what most papers mean by a CpG island.
• Takai and Jones (PNAS, 2002): 500 bp or more, GC content of 55% or more, and observed/expected CpG of 0.65 or more.

Why there are two, and a third in the genome browser

Takai and Jones tightened all three thresholds after scanning the complete sequences of human chromosomes 21 and 22. Their stricter regions were more likely to sit at the 5' ends of genes, and the definition excluded most Alu repeats — short, GC-rich, CpG-rich retrotransposon copies, of which the human genome carries more than a million.

UCSC's CpG-island track takes a third route: 1987-style thresholds (GC of 50% or more, longer than 200 bp, observed/expected above 0.6) applied to sequence with repeats masked out, which removes most Alus before they can be called at all.

Worked example 1: 100% GC and not an island

The sequence GGGGCCCC repeated 50 times is 400 bp of nothing but G and C, so its GC content is 100%. It contains 49 CpGs — one at each junction where a CCCC run meets the next GGGG — against an expectation of 200 × 200 ÷ 400 = 100. Its observed/expected ratio is 0.49, and it fails both definitions.

GC repeated 200 times is also 400 bp at 100% GC, but every second dinucleotide is a CpG: 199 of them, an observed/expected ratio of 1.99. It passes Gardiner-Garden and Frommer's criteria easily, and fails Takai and Jones' only because it is shorter than 500 bp.

Same length, same composition, opposite verdicts under the 1987 criteria. The arrangement of the bases decides, which is exactly the thing GC content cannot see.

Worked example 2: an Alu repeat

The AluY consensus sequence (Dfam family DF000000002, 311 bp including its poly-A tail) is 57.2% GC and contains 25 CpGs against 25.2 expected — an observed/expected ratio of 0.99. Under Gardiner-Garden and Frommer's criteria the whole element is a CpG island. Under Takai and Jones' it is not, and only because of length: it clears the observed/expected threshold comfortably (0.99 against 0.65) and the GC threshold narrowly (57.2% against 55%), but is shorter than 500 bp.

Genomic Alu copies score lower than the consensus, because their methylated CpGs have been decaying since each copy inserted. Young elements can still pass the 1987 test, and with more than a million of them in the genome, that is the problem the 2002 criteria and UCSC's repeat masking each solve in their own way.

Worked example 3: the GAPDH promoter

A 5 kb window of human chromosome 12 around the GAPDH promoter (NC_000012.12:6,532,001–6,537,000, GRCh38) is 60.8% GC overall. SeqBench applies both sets of thresholds with the same scan — Takai and Jones' 200 bp window and their rule of joining nearby passing regions — so the two results differ only in their thresholds. Both find a single island, and both islands span the gene's annotated start at 6,534,517: 1,892 bp under the 1987 thresholds (6,533,945–6,535,836; 65.5% GC, observed/expected 0.71) and 1,550 bp under the 2002 thresholds (6,533,981–6,535,530; 66.1% GC, observed/expected 0.76).

The flanking sequence outside the 1987 island is 57.9% GC — far above the human genome's average of about 41% — but its observed/expected CpG ratio is 0.23. That contrast is the whole idea in one locus: what marks the island is not a high GC content but the CpGs that survived.

UCSC's CpG-island track places its island at 6,534,096–6,535,441: 1,346 bp, 123 CpGs, observed/expected 0.8. Three results, one island, three sets of boundaries. They agree completely that the GAPDH promoter lies in a CpG island and disagree by a few hundred bases about where it stops: the two published definitions because their thresholds differ, and UCSC because it grows the region by a different rule, keeping the highest-scoring segment rather than merging windows. For 'is this promoter in an island?' any of them will do; for 'is this particular CpG inside the island?' near an edge, say which definition you used.

Using island boundaries in methylation work

Illumina's methylation array manifests label each probe by where it sits relative to an island: in the island, on a shore (up to 2 kb away) or on a shelf (2–4 kb away), prefixed N or S for the upstream or downstream side. Shores are where much of the tissue-specific methylation variation turns up, so knowing where an island ends is often more useful than knowing that it exists.

For bisulfite PCR, the island is where CpGs are densest, and assays are designed around them: bisulfite-sequencing primers are usually placed to avoid CpG sites so they amplify regardless of methylation, while methylation-specific PCR places CpGs in the primers deliberately. Either way, the first step is a map of where the island is.

Mistakes that change the answer

• Using GC content as a stand-in for CpG density. The GGGGCCCC example is 100% GC with an observed/expected ratio of 0.49.
• Comparing island calls from different tools as though they used one definition. The 1987 and 2002 criteria, and UCSC's masked scan, put the edges of the same GAPDH island hundreds of bases apart.
• Scanning sequence with repeats left in and reading every young Alu as an island under the 1987 criteria.
• Counting CpGs on both strands. The dinucleotide is palindromic, so each one is already counted once on either strand.
• Forgetting that the window and merge rules move boundaries. Takai and Jones scanned 200 bp windows and joined passing regions less than 100 bp apart, while Gardiner-Garden and Frommer's 1987 paper used a 100 bp moving window with no merge step; the same thresholds applied with different scan settings put the edges in different places.

Related references

Related tools

Related guides

• How to Fetch a Sequence by Accession Number →
• Identify Unknown Sequence: DNA, RNA or Protein →
• How to Make 0.5 M EDTA, 1 M Tris and Other Lab Stocks →