k-mer Frequency

The k-mer frequency distribution maps each possible k-length subsequence to its occurrence count in a genome. For a given k, there are 4^k possible k-mers (e.g., 4^4 = 256 for tetramers). Real genomes exhibit a highly skewed distribution: a small number of k-mers appear at extreme frequencies due to repetitive elements and functional constraints, while most occur at moderate levels. Synthetic data generators must reproduce this long-tailed distribution to avoid introducing artificial biases that would flag the data as non-biological. Key metrics include:

• Peak frequency: The most common k-mer count, typically driven by Alu repeats in human genomes
• Zero-frequency k-mers: Sequences absent due to biological constraints (e.g., restriction sites)
• GC-content correlation: k-mer frequency strongly correlates with local GC percentage

As k increases, the fraction of unique k-mers in a genome rises sharply. For the human genome (3.2 Gbp):

• At k=16, approximately 85% of observed k-mers are unique
• At k=20, over 99% are unique
• At k=30, essentially all observed k-mers appear exactly once

This uniqueness threshold is a critical property that synthetic generators must preserve. A model that produces too many repeated long k-mers indicates mode collapse or insufficient diversity. Conversely, failing to reproduce known repeats (e.g., LINE-1 elements) signals that the generator has not learned repetitive genome architecture. The transition point where k-mers become unique is a direct function of genome size and complexity.

The k-mer spectrum of sequencing errors produces a distinct low-frequency shoulder in the distribution. True genomic k-mers form a high-coverage peak, while sequencing errors generate a secondary peak at very low frequencies (typically 1-3x coverage). This bimodal distribution enables:

• Error correction: Distinguishing true variants from sequencing artifacts
• Genome size estimation: The area under the main peak estimates genome size
• Heterozygosity detection: A split main peak indicates heterozygous diploid genomes

Synthetic read generators must simulate this error-induced low-frequency tail to produce realistic FASTQ files. Omitting this property creates synthetic data that appears unrealistically clean and fails to stress-test downstream variant callers.

DNA is double-stranded, so a k-mer and its reverse complement represent the same biological entity. Canonical k-mer encoding collapses each k-mer pair into a single representative (lexicographically smaller of the two), reducing the feature space from 4^k to approximately 4^k/2. This encoding is essential for:

• Strand-agnostic analysis: Avoiding double-counting in frequency spectra
• Dimensionality reduction: Halving the input space for machine learning models
• Symmetry enforcement: Ensuring synthetic generators produce strand-symmetric outputs

Failure to canonicalize produces strand-biased spectra where forward-strand k-mers have different frequencies than reverse-strand k-mers—a clear artifact that real genomes do not exhibit.

Shannon entropy of the k-mer spectrum quantifies the information content and sequence complexity of a genome. High-entropy spectra indicate diverse, complex sequences; low-entropy spectra suggest repetitive or low-complexity regions. Key measures include:

• Linguistic sequence complexity: The ratio of observed unique k-mers to the maximum possible
• Conditional entropy: The predictability of the next base given the preceding (k-1)-mer
• Relative entropy (KL divergence): Comparing synthetic vs. real k-mer distributions

Synthetic generators with mode collapse produce artificially low-entropy spectra. Generators that introduce random noise produce artificially high-entropy spectra. The Frechet Genomic Distance uses these entropy-derived features to quantify the statistical fidelity of synthetic genomes.

When analyzing population cohorts, the aggregate k-mer spectrum reveals population structure beyond individual genomes. Key properties include:

• Private k-mers: Sequences unique to specific individuals, driven by rare variants
• Shared k-mers: Sequences common across the population, reflecting conserved functional elements
• Population-specific enrichment: k-mers enriched in specific ancestry groups

Synthetic population generators must preserve linkage disequilibrium patterns and allele frequency spectra to produce realistic cohort-level k-mer distributions. Differential privacy mechanisms add calibrated noise to these spectra, trading off population fidelity for privacy guarantees controlled by the epsilon parameter.

A learned, dense vector representation that maps discrete nucleotide bases (A, C, G, T) into a continuous space. Unlike simple one-hot encoding, embeddings capture biochemical similarities—for instance, placing purines (A, G) closer together than pyrimidines (C, T). This is critical for generative models because it allows the network to understand that a mutation from A to G is less biochemically severe than A to C. The quality of the embedding directly influences how well a model preserves k-mer frequency distributions in synthetic output.

A binary vector representation where each nucleotide base is mapped to a unique index position (e.g., A = [1,0,0,0], C = [0,1,0,0]). This serves as the simplest, unambiguous input format for synthetic genomic data generators. While computationally straightforward, it treats all mutations as equidistant, lacking the biochemical nuance of learned embeddings. When validating synthetic data, one-hot encoded k-mers provide a direct, interpretable check on whether the generator has preserved the exact k-mer frequency spectrum of the training data.

The proportion of guanine-cytosine bases in a DNA sequence, a key genomic signature that synthetic data generators must control. Real genomes exhibit species-specific GC content (e.g., ~41% in human). If a generative model introduces a shift—say, artificially inflating GC-rich k-mers—it distorts the k-mer frequency distribution and creates non-biological sequences. This bias can cause downstream tools like aligners or variant callers to fail. Monitoring GC content is a first-line sanity check for synthetic genomic data quality.

The ability of a generative model to accurately reproduce functional sequence patterns—such as transcription factor binding sites (e.g., TATA box, CTCF motifs)—in synthetic genomic data. These motifs represent specific, high-frequency k-mers that are evolutionarily conserved. A generator that fails to preserve motif-level k-mer frequencies produces sequences that look statistically plausible at the nucleotide level but are biologically inert. Motif preservation is evaluated by scanning synthetic sequences with position weight matrices (PWMs) and comparing hit rates to real genomic backgrounds.

A metric for evaluating synthetic genomic data quality by comparing the distribution of generated sequences to real sequences in a feature space. It extracts k-mer frequency vectors from both real and synthetic datasets, then computes the Frechet distance between the two multivariate Gaussian distributions fitted to these vectors. A lower distance indicates that the generator has faithfully reproduced the k-mer spectrum. This metric is the genomic analog of the Frechet Inception Distance (FID) used to evaluate image synthesis in computer vision.

A technique that trains a classifier to distinguish between real and synthetic genomic data based on their k-mer frequency profiles. If the classifier performs no better than random chance (AUC ≈ 0.5), the generator is considered high-quality—its k-mer distribution is indistinguishable from reality. This method is particularly sensitive to subtle distributional shifts that summary statistics might miss. It is commonly used during GAN training to detect mode collapse, where the generator fails to capture the full diversity of the real k-mer spectrum.