Kraken2

Kraken2 maps each k-mer (default k=35) from a sequencing read to a taxonomic identifier using a probabilistic hash table stored in memory. Unlike alignment-based tools, it performs exact-match queries against a pre-built database where every k-mer is associated with the lowest common ancestor (LCA) of all reference genomes containing that k-mer. This approach eliminates the computational overhead of base-by-base alignment while maintaining high sensitivity for species-level identification.

Kraken2 employs a minimizer technique to dramatically reduce database size. Instead of storing every k-mer, it selects the lexicographically smallest ℓ-mer (where ℓ < k) within each k-mer window. This reduces the database to approximately 8-10 GB for a complete RefSeq bacterial, archaeal, and viral collection—enabling classification on standard workstations without specialized hardware. The minimizer approach preserves sensitivity while achieving a 10-100x reduction in memory footprint compared to Kraken1.

When a read's k-mers map to multiple taxa, Kraken2 applies a conservative LCA algorithm to assign the read to the deepest taxonomic node shared by all matching references. This minimizes false-positive classifications by avoiding overconfident species-level calls when k-mers are shared across genera. The LCA approach is particularly effective for metagenomic samples containing closely related organisms or novel strains not present in the reference database.

Kraken2 computes a confidence score (0-1) for each classification by evaluating the proportion of k-mers supporting the assigned taxon versus alternative assignments. A score of 0.0 indicates low confidence (many k-mers unclassified or ambiguous), while 1.0 indicates high confidence (nearly all k-mers uniquely support the assignment). Users can apply a confidence threshold (typically 0.1-0.2) to filter spurious classifications, dramatically reducing false-positive rates in clinical and environmental samples.

Kraken2 is paired with Bracken (Bayesian Reestimation of Abundance after Classification), a companion tool that redistributes reads assigned to higher taxonomic levels down to species or genus level using a statistical model. Bracken estimates true abundance by accounting for reads that Kraken2 conservatively assigned to internal nodes due to shared k-mers. This yields species-level abundance profiles suitable for differential abundance analysis and ecological comparisons.

Kraken2 processes millions of reads per minute on a single CPU core by loading the entire database into RAM and performing constant-time hash lookups. It supports FASTQ and FASTA input formats and can classify both single-end and paired-end reads. The streaming architecture enables integration into real-time metagenomic pipelines for pathogen surveillance, outbreak monitoring, and clinical diagnostics where turnaround time is critical.

The core taxonomic assignment strategy used by Kraken2. When a query k-mer maps to multiple reference genomes, the algorithm assigns the read to the deepest node in the taxonomic tree that is a common ancestor of all matching genomes. This conservative approach minimizes false-positive species-level calls by defaulting to a higher rank (e.g., genus or family) when exact matches are ambiguous, ensuring high precision in metagenomic profiling.

The frequency distribution of all nucleotide subsequences of fixed length k within a sequencing read or genome. Kraken2 uses k=35 by default, mapping each k-mer to a taxonomic identifier via a compact hash table. The k-mer spectrum serves as the fundamental compositional feature for assembly-free classification, with the choice of k representing a trade-off between sensitivity (shorter k-mers) and specificity (longer k-mers).

Locality-sensitive hashing techniques that estimate Jaccard similarity between k-mer sets using compact sketches. While Kraken2 uses exact k-mer matching, tools like Sourmash employ FracMinHash for scalable genomic distance estimation. FracMinHash selects a fraction of k-mer hash values below a defined abundance threshold, providing robust containment index and Average Nucleotide Identity (ANI) estimates without full database construction.

A companion tool to Kraken2 that statistically reestimates species-level abundances from Kraken2's read-level classifications. While Kraken2 reports the number of reads assigned to each taxon, Bracken uses a Bayesian model to redistribute reads assigned to higher taxonomic ranks down to the species level, accounting for the fact that reads from the same genome may be classified at different ranks due to shared k-mers.

The process of grouping assembled contigs into Metagenome-Assembled Genomes (MAGs) based on sequence composition and abundance patterns. Kraken2's taxonomic labels can guide binning algorithms like MetaBAT2 and CONCOCT by providing initial taxonomic anchors. Binning complements classification by reconstructing draft genomes of uncultivated organisms, assessed for completeness and contamination using tools like CheckM.

The Comprehensive Antibiotic Resistance Database, a curated resource for antimicrobial resistance gene detection. Kraken2 can be paired with a custom database built from CARD sequences to perform AMR prediction directly from metagenomic reads. This enables rapid identification of known resistance determinants without assembly, critical for clinical surveillance and outbreak investigation workflows.