ToolsAbout

CpG island finder

Scan DNA for CpG islands by GC content and observed/expected CpG ratio.

CpG islands
How are CpG islands predicted?
A window (200 bp by default) slides along the sequence. For each position the tool counts C, G and the dinucleotide CG, then tests two things: the GC content and the ratio of observed to expected CpG. Windows that pass both tests and overlap are merged into one island. Islands shorter than the minimum length are dropped, and GC content and Obs/Exp are recalculated over each whole island for the table.
Obs/Exp CpG = (CpG count × N) / (C count × G count)
GC % = 100 × (C + G) / N
N is the number of bases in the window or island. The defaults are the criteria of Gardiner-Garden and Frommer (1987): length ≥ 200 bp, GC ≥ 50 %, Obs/Exp ≥ 0.6. For the GC content of a whole sequence use the GC content calculator.
Worked example
The example is a 620 bp sequence with an AT-rich flank on each side of a GC-rich core. With the default settings the tool reports one island from position 35 to 561, 527 bp long. It contains 154 C, 144 G and 54 CpG, so GC = 100 × 298 / 527 = 56.5 % and Obs/Exp = (54 × 527) / (154 × 144) = 1.28. The GC-rich core itself is bases 151–470: an island reaches into the flanks as far as a window that still passes both tests, so borders are accurate to a fraction of the window size.
What do the settings change?
  • Window: the stretch over which GC and Obs/Exp are measured. Short windows are noisy because the expected CpG count becomes small.
  • Step: how far the window moves each time. Step 1 gives the most exact borders; a larger step is only a speed setting and is rarely needed.
  • Min GC % and Min Obs/Exp: the two tests every window must pass.
  • Min length: merged islands shorter than this are left out.
Positions are 1-based and inclusive. N and other ambiguity codes count towards the length but not as C or G. Several FASTA records are scanned one by one.
Frequently asked questions
What is a CpG island?
A stretch of DNA, usually 200 bp to a few kb long, where the dinucleotide CG occurs about as often as the base composition predicts. Elsewhere in vertebrate genomes CG is depleted to roughly a fifth of the expected number, because methylated cytosine in CpG mutates to thymine. Islands sit at the promoters of about 60–70 % of human genes and are normally unmethylated.
What are the Gardiner-Garden and Frommer criteria?
Length of at least 200 bp, GC content of at least 50 % and an observed/expected CpG ratio of at least 0.6 (Gardiner-Garden and Frommer, J Mol Biol 1987). These are the defaults here.
What are the stricter Takai and Jones criteria?
Length of at least 500 bp, GC content of at least 55 % and observed/expected CpG of at least 0.65 (Takai and Jones, PNAS 2002). They exclude most Alu repeats, which often pass the 1987 criteria. Set window and minimum length to 500, GC to 55 and Obs/Exp to 0.65 to apply them.
How is the observed/expected CpG ratio calculated?
Obs/Exp = (number of CpG × N) / (number of C × number of G), where N is the length of the window or island. A value of 1 means CG occurs as often as chance predicts from the C and G counts.
Why can an island have values slightly below the thresholds?
Every window inside an island passes the thresholds, but the GC content and Obs/Exp shown in the table are recalculated over the whole merged island. Because Obs/Exp is not a simple average, the island value can differ a little from the window values.