Sequence Complexity (SEQ_CX)¶

This INFO field captures 11 sequence complexity features distilled from four raw measurement modules (homopolymer runs, Shannon entropy, LongdustQ k-mer concentration, tandem repeat motifs). Emitted as the SEQ_CX INFO tag in every VCF record.

All 11 features are computed from assembled haplotype sequences and are completely coverage-invariant: they measure properties of the genomic sequence itself, not properties of the reads. A poly-A run is 12bp long regardless of whether 20 or 2000 reads cover it.

INFO tag: SEQ_CX (11 comma-separated values)

The Four Measurement Modules¶

The 11 distilled features are built from four raw measurement modules, each capturing a different aspect of sequence complexity at different biological scales. Understanding these building blocks makes the distilled features intuitive.

1. Homopolymer Run (HRun)¶

The longest consecutive stretch of identical bases in a DNA window.

What it measures: The length of the single longest same-base run — e.g., AATTTTTTTTGC has an HRun of 8 (the T run).

Why it matters: Homopolymers are the #1 source of Illumina INDEL errors. During sequencing, the polymerase can "stutter" on long same-base runs, adding or dropping 1–2 bases. A 12bp poly-A run produces ~10× more false 1bp insertions than a random 12bp sequence. HRun directly quantifies this stutter risk.

Computation: Single O(L) scan counting consecutive identical characters. Returns 0 for empty sequences, 1 for single-base sequences. No parameters.

2. Shannon Entropy¶

Base frequency diversity in a DNA window, measured in bits (0.0–2.0).

What it measures: How evenly the four DNA bases (A, C, G, T) are distributed. The formula H = −Σ pᵢ log₂(pᵢ) counts the information content of the base composition.

Interpretation:

• 0.0 = all one base (e.g., AAAAAAA) — maximally repetitive
• 1.0 = two equally frequent bases (e.g., alternating ACACAC)
• 2.0 = perfectly balanced ACGT — maximally diverse

Why it matters: Low-entropy regions are harder to align uniquely because many positions look identical to the aligner. Variants called in low-entropy windows are more likely to be mismapping artifacts. Shannon entropy captures this alignment ambiguity risk.

Computation: O(L) base counting, then the entropy formula. No parameters.

3. LongdustQ (k-mer Concentration)¶

A k-mer-based complexity score that measures how repetitive a sequence is by comparing the observed k-mer count distribution against a random (Poisson) null model.

What it measures: Whether any short DNA words (k-mers) appear more often than expected by chance. In random DNA, each k-mer appears roughly equally. In repetitive DNA, a few k-mers dominate — and LongdustQ quantifies this concentration.

Interpretation:

Why it matters: LongdustQ detects repetitive structures that HRun and entropy miss — microsatellites, tandem repeats, and satellite DNA that cause systematic assembly errors. A dinucleotide repeat (CACACACA) has high entropy (1.0) and short HRun (1), but LongdustQ correctly flags it as repetitive because two k-mers dominate the distribution.

Provenance: Adapted from Li, H. "Finding low-complexity filter with longdust" (2025). Lancet2 extracts just the Q(x) scoring formula and applies it directly to known subsequences around variants — no sliding-window scanning is needed because variant locations are already known from graph assembly. Two k-mer sizes are used: k=4 for local ±50bp flanks (sensitive to short repeats) and k=7 for full haplotype scoring (sensitive to macro-scale satellite structure).

GC-bias correction: LongdustQ supports an optional GC-bias correction (default: human genome GC=0.41) to prevent AT-rich regions from being falsely scored as repetitive. See GC-Bias Correction.

4. Tandem Repeat (TR) Motif Detection¶

Finds exact and approximate short tandem repeats (period 1–6) in a flanking window around the variant.

What it measures: The presence, proximity, purity, and period of the nearest tandem repeat to the variant site. Searches for motifs that repeat ≥2.5× (exact) or ≥3.0× (approximate, allowing 1 edit per repeat unit).

Why it matters: Repeat proximity is the single strongest predictor of Illumina INDEL artifacts. A 1bp insertion adjacent to a (CA)₁₀ dinucleotide repeat is overwhelmingly likely to be polymerase stutter, not a real mutation. TR detection directly identifies this pattern and flags the canonical stutter signature.

Key outputs:

• Distance: bp distance from variant to nearest TR (0 = overlapping)
• Purity: fraction of the repeat span that is error-free (1.0 = perfect)
• Period: motif length (1=homopolymer, 2=dinucleotide, ..., 6=hexanucleotide)
• Stutter flag: 1 if INDEL length ≤ period AND adjacent to TR (≤1bp)

Computation: O(L × P) where L = window length and P = max period (6). See Motif Detection Parameters for defaults.

Design: Context vs. Perturbation Paradigm¶

Raw sequence complexity metrics at multiple scales (±5/10/20/50/100bp, full haplotype) are highly correlated — HRun at ±5bp, ±10bp, and ±20bp are ~0.95 correlated. Additive ML models — such as EBMs (Explainable Boosting Machines, which learn one feature's effect at a time, then combine them) and GAMs (Generalized Additive Models) — split importance weight across correlated features, producing noisy shape functions instead of confident ones.

The distillation separates metrics into three groups that are independent between groups (Context, Delta, TR Motif measure fundamentally different things), while features within each group are deliberately chosen at non-overlapping scales to minimize correlation:

1. Context (REF-only): "How brittle is the genome here, regardless of the variant?"
2. Deltas (ALT−REF): "How did the variant alter the local complexity?"
3. TR Motif (ALT-only): "What is the repeat environment of the final allele?"

This enables the model to learn rescue logic: even if context is dangerous (high ContextHRun), a negative DeltaHRun (variant broke the homopolymer) rescues the call.

Feature Layout¶

A. Context (4 features — strictly REF)¶

B. Deltas (3 features — ALT minus REF)¶

C. TR Motif (4 features — strictly ALT ±50bp)¶

Transform Details¶

Sentinel-Safe TR Features¶

Raw motif detection uses dist_to_nearest_tr = −1 as a sentinel for "no TR found." This breaks EBM numerical binning because −1 sorts between −2 and 0, making "no TR" appear closer to "sitting on the repeat" than a distance of 10.

The TrAffinity inverse transform (1/(1+dist)) maps the sentinel into the continuous [0, 1] range with 0 = no TR and 1 = overlapping TR. All other TR features are zeroed when no TR is found.

Log₁p Squashing¶

LongdustQ reaches 4.0+ in telomeric and pericentromeric regions, creating extreme heavy tails. The log₁p(max(0, x)) transform compresses these tails while preserving monotonicity and differentiability at zero.

Multi-Haplotype and Multi-Allelic Merging¶

When a variant appears on multiple ALT haplotypes, the scorer produces one SequenceComplexity per haplotype and merges via element-wise max (pessimistic worst-case). Multi-allelic variants at the same locus are similarly merged.

GC-Bias Correction¶

LongdustQ uses a null model for expected k-mer frequencies. By default, it assumes GC fraction = 0.41 (human genome-wide average). Without correction, AT-rich regions are scored as artificially "repetitive."

--genome-gc-bias 0.41    # Human genome (default)
--genome-gc-bias 0.5     # Uniform model (no correction)
--genome-gc-bias 0.42    # Mouse genome

Coverage Stability¶

All 11 SEQ_CX features are perfectly coverage-invariant because they are computed from assembled haplotype sequences, not from read-level metrics. The scoring pipeline works as follows:

1. The de Bruijn graph assembler produces haplotype sequences (contigs) from the reads in each window.
2. The sequence complexity scorer operates on these haplotype strings.
3. Homopolymer runs, Shannon entropy, LongdustQ, and TR motif detection all operate on the sequence characters, not on read counts or qualities.

The only indirect coverage dependency is that the assembler itself requires sufficient coverage to produce correct haplotypes. Below ~10×, the assembler may fail to construct the ALT haplotype, in which case no variant is called and no SEQ_CX is computed. Above ~15×, the assembled haplotypes are identical regardless of depth, and all 11 features are deterministic:

Motif Detection Parameters¶

The tandem repeat motif detector scans for exact and approximate repeats. Default configuration is optimized for somatic variant calling on Illumina short reads: