Canonical K-mers

The canonical form is defined by comparing the original k-mer string against its reverse complement and selecting the one that comes first alphabetically. For example, given the 3-mer ACG, its reverse complement is CGT. Since ACG < CGT lexicographically, ACG is the canonical representation. This deterministic rule ensures that both ACG and its reverse complement CGT map to the same feature, eliminating redundant strand-specific entries.

Without canonicalization, a k-mer vocabulary would contain up to 4^k distinct features. By collapsing each k-mer and its reverse complement into a single feature, the feature space is approximately halved. For k=7, the raw vocabulary of 16,384 possible heptamers reduces to roughly 8,000–8,500 canonical features, significantly decreasing model sparsity and computational overhead without losing biological information.

Sequencing reads originate from either the forward or reverse strand of DNA, but the underlying biological entity—a binding site or regulatory element—is strand-agnostic. Canonical k-mers enforce strand symmetry by design: a transcription factor binding motif and its reverse complement produce identical feature vectors. This prevents the model from learning spurious strand-specific patterns and improves generalization.

A k-mer that equals its own reverse complement is a biological palindrome (e.g., GAATTC for k=6). In these cases, the canonical form is trivially the k-mer itself, as both strands produce the same sequence. Palindromic k-mers are often enriched at restriction enzyme cut sites and dimeric transcription factor binding motifs, making their unambiguous representation particularly important for regulatory genomics models.

Canonical k-mers are commonly used as keys in hash tables for k-mer counting and feature indexing. By normalizing to the canonical form before hashing, tools like Jellyfish and KMC ensure that a k-mer and its reverse complement map to the same bucket. This property is essential for memory-efficient De Bruijn graph construction in genome assembly and for consistent feature mapping in machine learning pipelines.

Models like DNABERT tokenize sequences into overlapping k-mers and rely on a fixed vocabulary. Canonicalization ensures that the vocabulary size remains manageable and that semantically equivalent reverse-complement tokens share the same embedding vector. This pre-processing step is typically applied during vocabulary construction, so the tokenizer directly maps both ACG and CGT to the same integer ID before training.

A sequence vectorization technique that decomposes a nucleotide string into all possible substrings of length k, then maps each k-mer to a unique numerical identifier or frequency vector. This transforms raw biological sequences into a fixed-dimensional feature space suitable for classical machine learning algorithms. The choice of k directly controls the trade-off between sequence specificity and feature sparsity—smaller k values capture local nucleotide composition, while larger k values resolve longer functional motifs at the cost of exponential vocabulary growth.

A data augmentation and representation strategy that explicitly accounts for the double-stranded nature of DNA by ensuring a sequence and its reverse complement map to identical or equivalent embedding vectors. Without this, a model would treat the forward strand and its reverse complement as independent features, doubling the effective vocabulary and diluting statistical power. Canonical k-mers implement this principle directly by collapsing each k-mer and its reverse complement into a single, unambiguous feature using a deterministic selection rule.

A sparse binary representation where each nucleotide (A, C, G, T) is mapped to a distinct orthogonal basis vector, creating a 4-channel matrix that serves as a raw input for convolutional neural networks. Unlike k-mer encoding, one-hot encoding preserves the exact positional information of every base, allowing convolutional filters to learn motifs directly from the raw sequence. This representation is the standard input format for architectures like Enformer and BPNet, which learn their own internal motif representations rather than relying on pre-defined k-mer features.

The foundational genomic language model DNABERT adapts the BERT architecture by tokenizing DNA sequences into overlapping k-mers and pre-training via masked language modeling on the human reference genome. Its tokenizer uses a sliding window of k=3 to 6 to generate context-aware nucleotide embeddings, where each k-mer is treated as a word in the genomic vocabulary. The resulting embeddings capture the regulatory syntax and functional grammar of non-coding DNA, enabling transfer learning to downstream tasks like promoter prediction and splice site detection.

A tokenization strategy that segments an mRNA transcript or coding DNA sequence into non-overlapping triplets of nucleotides, directly aligning the vocabulary with the fundamental functional units of the genetic code. This biologically-informed approach reduces the sequence length by a factor of three while preserving the translational reading frame. When combined with canonical k-mer logic, codon tokenization ensures that synonymous codons and their reverse complements are handled consistently, making it particularly effective for protein-coding gene prediction and variant effect modeling.

The IUPAC nucleotide notation (e.g., N for any base, R for purine, Y for pyrimidine) used to represent positions of uncertainty or natural variation within a consensus sequence. These degenerate characters require specialized embedding strategies to handle input sequences that cannot be cleanly decomposed into standard k-mers. When applying canonical k-mer selection to sequences containing ambiguity codes, the reverse complement mapping must account for the full IUPAC complementarity rules—for example, R (A or G) pairs with Y (C or T)—to maintain strand-invariant representation.