The G4Hunter Algorithm

This page provides a detailed explanation of the G4Hunter algorithm as described by Bedrat, Lacroix & Mergny (2016) [1]. Understanding the algorithm helps interpret g4hunterpy3 output and choose appropriate parameters.

What Are G-Quadruplexes?

G-quadruplexes (G4) are four-stranded nucleic acid structures formed by guanine-rich sequences. Their building block is the guanine quartet (G-quartet) — a planar association of four guanines held together by Hoogsteen hydrogen bonds. Two or more stacked G-quartets, stabilized by coordinated cations (typically K⁺), form a G-quadruplex.

G4 structures have attracted significant biological interest because they:

• Are involved in telomere biology and chromosome end-capping.

• Play roles in transcription regulation (e.g., the MYC promoter G4).

• Affect DNA replication, acting as replication barriers.

• Are implicated in genomic instability — mitochondrial DNA deletion breakpoints map near G4-forming regions.

• Serve as drug targets for anticancer and antiviral therapeutics.

• Are found enriched in gene promoters, particularly proto-oncogenes.

Why G4Hunter?

Prior to G4Hunter, the dominant approach for predicting G4-forming sequences was Quadparser, which searched for patterns matching the motif G_n N_m G_n N_o G_n N_p G_n, requiring fixed-length G-runs and constrained loop sizes. While useful, Quadparser has important limitations:

• False negatives: sequences that do form G4 structures experimentally but lack the canonical pattern are missed. Some G4 structures contain interrupted G-runs, bulges, or non-standard loop lengths.

• False positives: some sequences matching the consensus fail to form G4 structures in vitro because competing duplex formation (particularly in GC-alternating regions) is more favorable.

• Binary output: Quadparser provides only a yes/no answer, without a quantitative measure of G4-forming propensity.

G4Hunter addresses all three issues by providing a continuous score based on two sequence properties that are correlated with G4 formation:

1. G-richness — the density and clustering of guanines.

2. G-skewness — the asymmetry between G and C content on a given strand.

Per-Base Scoring

The foundation of G4Hunter is a per-base scoring scheme that assigns every nucleotide in a sequence a score between -4 and +4:

Base	Score	Rationale
A, T (U)	0	Neutral — do not contribute to or compete with G4 formation.
G in a single-G context	+1	Weak G-richness signal.
Each G in a GG run	+2	Moderate G-richness.
Each G in a GGG run	+3	Strong G-richness — the classic G-tract.
Each G in a run of 4+ Gs	+4 (capped)	Very strong G-richness (capped at 4).
C	Negative (same magnitude rules as G)	C-runs oppose G4 formation on the scored strand, reflecting competition from stable Watson-Crick duplexes.

Worked example:

Sequence:   G G G A T T C C G G G G A T
Run lengths: 3 3 3 - - - 2 2 4 4 4 4 - -
Scores:     +3+3+3  0 0-2-2+4+4+4+4  0 0

Key design properties:

• Runs of G are rewarded superlinearly. A GGG run scores 3 per base (total 9), far more than three isolated G’s (total 3). This captures the physical reality that consecutive guanines are needed to form G-quartets.

• C-runs receive negative scores. This simultaneously (a) penalizes GC-alternating regions where stable duplex formation competes with G4, and (b) allows scoring the complementary strand — a highly negative score on one strand implies a highly positive score (and G4 potential) on the reverse complement.

• Scores are capped at ±4. Runs longer than 4 still receive a score of 4 per base. This prevents extremely long G-tracts from dominating the score disproportionately.

Sliding Window Scoring (G4Hscore)

For genome-wide or sequence-level analysis, per-base scores are smoothed using a sliding window:

1. Fix a window size k (default: 25 nt).

2. Slide the window one base at a time across the sequence.

3. For each window position, compute the arithmetic mean of the per-base scores within that window. This mean is the G4Hscore for that window.

\[\text{G4Hscore}(i) = \frac{1}{k} \sum_{j=i}^{i+k-1} \text{score}(j)\]

The G4Hscore has several useful properties:

• It is centered on zero for random sequences, independent of GC content. This is because G and C contributions cancel in sequences without strand asymmetry.

• Positive values indicate G-rich regions with G4-forming potential on the scored (forward) strand.

• Negative values indicate C-rich regions, implying G4-forming potential on the complementary strand.

• The magnitude reflects G4 propensity strength.

Window size considerations

The default window size of 25 nt was chosen to match the typical length of experimentally characterized G-quadruplex-forming sequences (~26 nt mean in the Bedrat et al. reference dataset). The window size can be adjusted:

• Smaller windows (15–20 nt) increase sensitivity for short G4 motifs but may increase false positives, especially from single long G-runs that cannot form intramolecular G4 structures alone.

• Larger windows (30–100 nt) can identify broader regions where multiple G4 structures may form in tandem, which may be biologically relevant (e.g., for replication-associated DNA damage).

In g4hunterpy3, window size is set via window_size in the Python API or -w on the command line.

Thresholding and Hit Calling

Windows whose absolute G4Hscore exceeds a user-defined threshold are reported as candidate G4-forming regions. Taking the absolute value means that G4-forming potential on both strands is captured simultaneously.

\[|\text{G4Hscore}(i)| \geq \theta\]

The threshold $theta$ controls the trade-off between sensitivity and specificity:

Threshold	Precision	Recommended use
1.0	~73%	Most inclusive; recovers the maximum number of true G4-forming sequences but with a higher false positive rate. Miss rate ~6%.
1.2	~85%	Recommended compromise. Identifies many true G4 motifs while maintaining reasonable precision. Miss rate ~15%.
1.5	>90%	High-confidence predictions. Most sequences identified at this threshold are experimentally confirmed G4 formers.
1.75	~95%	Very stringent; near-zero false positives on the mitochondrial genome validation set. Best for applications requiring certainty (e.g., PCR primer or DNA origami design).
2.0	>98%	Most stringent; identifies only the strongest G4-forming sequences.

These precision estimates are based on experimental validation of 209 sequences from the human mitochondrial genome using six independent biophysical methods (CD, NMR, TDS, IDS, UV-melting, and thioflavin T fluorescence).

In g4hunterpy3, the threshold is set via threshold in the Python API or -s on the command line.

Region Merging

In a genome-wide scan, consecutive windows with scores above the threshold typically overlap extensively (windows shifted by 1 nt share k-1 bases). These overlapping windows are merged into contiguous regions:

1. Sort window hits by start position.

2. Merge windows whose start positions overlap with existing regions (i.e., next window start ≤ current region end - 1).

3. For each merged region, compute the region score as the mean of the per-base scores across the full region (not the mean of window scores).

4. Report the merged region with its coordinates, sequence, length, score, and number of contributing windows.

This produces a non-redundant set of candidate G4-forming sequences (G4FS) from the genome.

Scoring Both Strands Simultaneously

A key advantage of the G4Hunter scoring scheme is that both strands of a DNA duplex are scored simultaneously. Because C-runs receive negative scores while G-runs receive positive scores:

• A region with a positive G4Hscore has G4-forming potential on the forward (scored) strand.

• A region with a negative G4Hscore has G4-forming potential on the reverse complement strand (which is G-rich where the forward strand is C-rich).

This means a single pass through a sequence identifies G4 candidates on both strands. The --strand-agnostic option in the CLI takes absolute values of scores, treating both strands equally when plotting.

Comparison with Quadparser

Bedrat et al. systematically compared G4Hunter against Quadparser on a reference dataset of 392 sequences with experimentally known G4 formation status, and on 209 sequences from the human mitochondrial genome:

Method	True G4 found	False positive rate	Type
G4Hunter (θ=1.0)	281 / 298	10.6%	Continuous score
G4Hunter (θ=1.2)	252 / 298	6.4%	Continuous score
Quadparser (QP37)	196 / 298	1.1%	Binary
Quadparser (QP27)	More hits	Higher FPR	Binary

Key advantages of G4Hunter:

1. Fewer false negatives. G4Hunter identifies G4-forming sequences that lack the canonical (G₃₊N₁₋₇)₄ pattern — sequences with bulges, interrupted G-runs, or non-standard loop lengths.

2. Quantitative output. The continuous score allows ranking sequences by G4-forming propensity and tuning sensitivity via the threshold.

3. Both strands. A single scan reports candidates on both strands.

4. ROC AUC > 0.96 on the reference dataset, indicating excellent discriminating power.

Genomic Distribution of G4-Forming Sequences

Application of G4Hunter to 20 genomes revealed:

• Mammalian genomes are the most G4-rich, with ~2.5 G4FS per kb at θ=1.0 and ~0.5 per kb at θ=1.5 (window size 25).

• G4FS density decreases exponentially with increasing threshold, with similar exponential slopes across mammalian species.

• Promoter regions are significantly enriched in G4FS (2–5.5-fold) compared to genomic background, particularly for stable G4 motifs (θ ≥ 1.5).

• Enrichment is also observed in 5’ UTRs, first exons, and first exon/intron junctions, predominantly on the coding strand.

• The number of G4-forming sequences in the human genome is estimated to be 2–10 times higher than the ~376,000 previously reported using Quadparser.

Implementation in g4hunterpy3

g4hunterpy3 implements the G4Hunter algorithm as described above with the following pipeline:

1. base_scores() — assigns per-base scores using the run-length scoring scheme.

2. window_mean_scores() — computes sliding-window means using fast O(n) convolution.

3. find_window_hits() — identifies windows exceeding the absolute-score threshold.

4. merge_overlapping_windows() — merges overlapping hits into non-redundant regions with per-base-mean region scores.

The convenience function scan_sequence() runs all four steps, and scan_fasta() applies the pipeline to every record in a FASTA file.

References

[1]

Bedrat, A., Lacroix, L. & Mergny, J.-L. Re-evaluation of G-quadruplex propensity with G4Hunter. Nucleic Acids Res. 44, 1746–1759 (2016). doi:10.1093/nar/gkw006