Introduction

Centrifuger is an efficient taxonomic classification method that compares sequencing reads against a microbial genome or protein database. Given a set of reads and a reference database such as RefSeq, GTDB or NCBI nr, Centrifuger reports, for every read, which reference sequence it came from and the taxonomic identifier of that sequence. A companion program turns those per-read assignments into an abundance profile for the sample.

What distinguishes Centrifuger from other classifiers is how the reference database is stored. It implements a novel lossless compression method — the run-block compressed BWT — and combines it with other strategies for compacting the Ferragina–Manzini (FM) index. The database stays complete: no k-mers are dropped, no minimizers are subsampled, and no sequence is discarded. It simply takes much less room.

Why compression matters#

Comprehensive reference databases have grown faster than the memory of the machines used to search them. The 2023 RefSeq prokaryotic collection is roughly 140 billion nucleotides; NCBI core_nt and nr are larger still. Classifiers generally cope in one of two ways: they throw information away — subsampling k-mers, restricting matches to a fixed length — or they demand a very large machine.

Centrifuger takes the third route. Microbial genome databases sit at an intermediate level of repetitiveness: far more repetitive than a single genome, far less repetitive than a collection of thousands of strains of one species. Centrifuger introduces two compact data structures tuned for exactly that regime:

• Run-block compressed BWT (RBBWT) — a run-block compressed sequence representation that achieves sublinear storage for the BWT without sacrificing much time efficiency.
• Hybrid run-length compressed BWT — a complementary representation used where run-length coding pays off.

Together with other FM-index compaction strategies, this halves the memory footprint compared with other FM-index-based approaches, letting the 2023 RefSeq prokaryotic database be classified against in about 43 GB of memory.

Why lossless matters#

Compression is not only an engineering convenience here — it is what makes the accuracy possible.

Because the representation is lossless and the index is a true FM-index, Centrifuger can extend a match for as long as the read and the reference agree. There is no fixed k that truncates a match, and no sampled k-mer set that can miss one. That unconstrained match length, together with the complete database, is what gives Centrifuger significantly better sensitivity and precision for species- and genus-level classification.

What is in the distribution#

Centrifuger ships as a small set of command-line programs:

The usual workflow is a straight line through them:

reference FASTA + taxonomy  ──centrifuger-build──▶  index (*.cfr)
                                                       │
                       reads (FASTQ)  ──centrifuger────┘──▶  classification.tsv
                                                                    │
                                        centrifuger-quant ──────────┘──▶  report.tsv

If a published index already covers your reference set, you can skip the first step entirely — see pre-built indexes.

What Centrifuger handles#

• Read layouts — paired-end (-1/-2), single-end (-u), interleaved (-i), and many samples at once through a sample sheet. The option follows the layout of the data, not its read length.
• Read lengths — short-read and long-read data, with no separate mode for either.
• Assay types — bulk sequencing as well as barcoded single-cell data, including 10x Genomics and combinatorial barcoding such as SHARE-seq.
• Sequence alphabets — nucleotide databases, and protein databases through centrifuger-build --protein.
• Report formats — the Centrifuge report format plus MetaPhlAn, CAMI and Kraken-report output from centrifuger-quant.

Relationship to Centrifuge#

Centrifuger follows the design and conventions of Centrifuge, from the same group, closely enough that the index-building procedure and the default report format will look familiar; the programs are renamed (centrifuge-download becomes centrifuger-download, and so on). The important differences are the compressed index, the unconstrained match length, and a few conveniences such as file-list input instead of concatenated FASTA. If you are coming from Centrifuge, building an index points out where the two diverge.

Where to go next#

• Installation — Bioconda, or make from source.
• Quick start — classify a sample end to end with the bundled example data.
• Command-line interface — every option of every program in one place.