CpG Island Detection

A genomic region must exceed a minimum length to qualify as a CpG island, filtering out short, stochastic CpG clusters that lack regulatory significance.

• Standard minimum: 200 base pairs (bp), as established by Gardiner-Garden and Frommer (1987)
• Extended minimum: 500 bp, used in more stringent definitions to reduce false positives
• Rationale: Short regions with high GC content occur frequently by chance; the length filter ensures only biologically meaningful clusters are retained
• Implementation: A sliding window scans the genome, and any window meeting all criteria is extended until the conditions fail

The proportion of guanine (G) and cytosine (C) nucleotides within the candidate window must exceed a defined threshold, reflecting the compositional bias of CpG islands.

• Standard threshold: Greater than 50% GC content
• Calculation: (Count of G + Count of C) / Total nucleotides in window
• Biological basis: CpG islands are predominantly located in GC-rich regions, often near housekeeping gene promoters
• Edge case: Some algorithms use a 55% threshold for higher specificity in organisms with globally high GC content

This ratio compares the actual frequency of CpG dinucleotides to the frequency expected if G and C were randomly distributed, serving as the most discriminative criterion.

• Formula: (Number of CpG × Length) / (Number of C × Number of G)
• Standard threshold: Observed/Expected ratio ≥ 0.60
• Interpretation: A ratio near 1.0 indicates no CpG depletion; values below 0.60 indicate the characteristic suppression seen in bulk DNA due to methylation-driven deamination
• Significance: This metric directly captures the evolutionary pressure that preserves CpG sites in regulatory regions while they are depleted elsewhere

The detection process employs a moving window that scans the genome incrementally, evaluating each candidate segment against all three criteria simultaneously.

• Window size: Typically 200 bp, shifting by 1 bp or a fixed step size
• Scoring: Each window receives a composite score based on GC content and CpG O/E ratio
• Merging: Adjacent windows that pass all thresholds are merged into contiguous CpG island annotations
• Optimization: Modern implementations use cumulative sums for O(1) window updates, enabling efficient whole-genome scanning without recomputing statistics from scratch

A refined definition proposed by Takai and Jones (2002) that applies stricter thresholds to exclude Alu repetitive elements and other non-regulatory GC-rich regions.

• Length: ≥ 500 bp
• GC content: ≥ 55%
• CpG O/E ratio: ≥ 0.65
• Advantage: Reduces false positives by approximately 75% compared to the original Gardiner-Garden criteria
• Application: Preferred in human genome studies where Alu elements, which are GC-rich but non-regulatory, would otherwise inflate CpG island counts

Beyond rule-based criteria, probabilistic frameworks like Hidden Markov Models (HMMs) classify genomic regions by learning transition probabilities between CpG island and non-island states.

• States: CpG island (+) and background (-) with distinct nucleotide emission probabilities
• Training: Baum-Welch algorithm on annotated genomes to learn state transitions
• Decoding: Viterbi algorithm assigns each nucleotide to the most likely state path
• Advantage: HMMs capture the gradual transition zones at CpG island boundaries, which rigid sliding windows often miss
• Tools: CpGIE and CpGProd implement HMM-based detection with species-specific parameterization

The core biological unit of detection. A CpG site is a cytosine followed by a guanine in the 5' to 3' direction. In vertebrates, 70-80% of CpGs are methylated, but those in CpG islands near promoters remain unmethylated, allowing gene expression. Detection algorithms exploit this statistical deviation from the genome-wide CpG depletion caused by the spontaneous deamination of methylated cytosine to thymine.

The original algorithmic definition. A CpG island is a region ≥200 base pairs with a G+C content >50% and an observed-to-expected CpG ratio >0.6. This heuristic remains the biological gold standard, though modern implementations often adjust thresholds for specific species or use statistical models to reduce false positives in GC-rich non-regulatory regions.

A refined definition to exclude Alu repetitive elements that mimic island properties. Criteria: length ≥500 bp, G+C content ≥55%, and an observed-to-expected CpG ratio ≥0.65. This higher stringency identifies islands more likely to be associated with housekeeping gene promoters and tumor suppressor genes, reducing false positives in repeat-masked genomes.

A probabilistic alternative to fixed-threshold methods. HMMs model the genome as a sequence of hidden states (e.g., CpG island vs. non-island) with distinct nucleotide emission probabilities. The Viterbi algorithm decodes the most likely path of states, identifying island boundaries with greater sensitivity. This approach captures transitional zones and short islands that rigid sliding-window methods miss.

The most direct computational implementation. A window of fixed size (e.g., 200 bp) shifts across the sequence in defined steps. At each position, the algorithm calculates:

• G+C percentage: (G + C) / window length
• Observed/Expected CpG ratio: (Number of CpGs × window length) / (Number of C × Number of G) Overlapping windows meeting thresholds are merged into contiguous island calls. This method is deterministic and highly interpretable.

A clinically significant downstream analysis. CIMP describes cancers exhibiting concerted hypermethylation of multiple CpG islands, silencing tumor suppressor genes. Detection involves clustering methylation array data to classify tumors as CIMP-high, CIMP-low, or CIMP-negative. This stratification has prognostic value in colorectal cancer and glioblastoma, guiding epigenetic therapy decisions.