Taxor: Fast and space-efficient taxonomic classification of long reads with hierarchical interleaved XOR filters
Ulrich, J. U., & Renard, B. Y. (2024). Fast and space-efficient taxonomic classification of long reads with hierarchical interleaved XOR filters. Genome Research, gr-278623. doi: 10.1101/gr.278623.123
Taxor is a taxonomic classification and profiling tool that efficiently classifies DNA sequences against large sets of genomic reference sequences. Taxor stores k-mers in an optimized hierarchical interleaved XOR filter (HIXF) index and combines k-mer similarity and genome coverage information for precise taxonomic classification and profiling. It features:
- Low false positive rates for k-mer matching
- NCBI taxonomy integration
- Open canonical syncmers as k-mer selection scheme for improved downsampling
- classification with binning and taxonomic profiling
- read reassignment EM algorithm for multi-matching reads
- advanced filtration of search results
- taxonomic and sequence abundance reports with genome size correction
Benchmarking results based on simulated and real long-read data sets demonstrate that Taxor enables more precise taxonomic classification and profiling of microbial populations while having a smaller memory footprint than other tools.
The easiest way is to install Taxor via Conda.
However, you can also build Taxor on your own using the following commands. Just make sure that you have installed CMake (>=3.16) and GCC (>= 10).
git clone https://github.com/JensUweUlrich/Taxor.git
cd Taxor
mkdir build
cd build
cmake ../src
cmake --build . --config Release
Users can easily build custom databases as described below or use the following pre-built database index files
| Kingdom | Source | Parameters | Size | Link | MD5 Hash |
|---|---|---|---|---|---|
| Viruses | Genbank Release 258 | k=22, s=12 | 373 MB | download | 0e8edd19a6314450f88f556dcf6b7c95 |
| Archaea, Bacteria, Fungi, Viruses | RefSeq Release 216 | k=22, s=12 | 9.9 GB | download | 768ee0320dcf41f5b15efafa028ba836 |
| Subcommand | Function |
|---|---|
| build | Construct HIXF index from fasta reference files |
| search | Search sequences against a database index |
| profile | Generate the taxonomic profile from search results |
taxor-build - Creates and HIXF index of a given set of fasta files
==================================================================
DESCRIPTION
Creates an HIXF index using either k-mers or syncmers
OPTIONS
Basic options:
-h, --help
Prints the help page.
-hh, --advanced-help
Prints the help page including advanced options.
--version
Prints the version information.
--copyright
Prints the copyright/license information.
--export-help (std::string)
Export the help page information. Value must be one of [html, man].
Main options:
--input-file (std::string)
tab-separated-value file containing taxonomy information and reference file names
--input-sequence-dir (std::string)
directory containing the fasta reference files Default: .
--output-filename (std::string)
A file name for the resulting index. Default: .
--kmer-size (signed 32 bit integer)
size of kmers used for index construction Default: 20. Value must be in range [1,30].
--syncmer-size (signed 32 bit integer)
size of syncmer used for index construction Default: 10. Value must be in range [1,26].
--threads (signed 32 bit integer)
The number of threads to use. Default: 1. Value must be in range [1,32].
--use-syncmer
enable using syncmers for smaller index size
input-file
This file contains all relevant information about the organisms in the database, which will be indexed. All values are tab-separated and the file should have following columns:
- Column 1: Assembly accession: the assembly accession.version reported in this field is a unique identifier for the set of sequences in this particular version of the genome assembly.
- Column 2: Taxonomy ID: the NCBI taxonomy identifier for the organism from which the genome assembly was derived. The NCBI Taxonomy Database is a curated classification and nomenclature for all of the organisms in the public sequence databases. The taxonomy record can be retrieved from the NCBI Taxonomy resource: https://www.ncbi.nlm.nih.gov/taxonomy/
- Column 3: FTP path: the path to the directory on the NCBI genomes FTP site from which data for this genome assembly can be downloaded
- Column 4: Organism name
- Column 5: Taxonomy string
- Column 6: Taxonomy ID string
A two-line example of such a file is provided below. You can easily create such a file by following the preprocessing steps described in the Usage section.
GCF_000002495.2 318829 https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/000/002/495/GCF_000002495.2_MG8 Pyricularia oryzae k__Eukaryota;p__Ascomycota;c__Sordariomycetes;o__Magnaporthales;f__Pyriculariaceae;g__Pyricularia;s__Pyricularia oryzae 2759;4890;147550;639021;2528436;48558;318829
GCF_000002515.2 28985 https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/000/002/515/GCF_000002515.2_ASM251v1 Kluyveromyces lactis k__Eukaryota;p__Ascomycota;c__Saccharomycetes;o__Saccharomycetales;f__Saccharomycetaceae;g__Kluyveromyces;s__Kluyveromyces lactis 2759;4890;4891;4892;4893;4910;28985
input-sequence-dir
Path to the directory containing fasta files (compressed) of organisms listed in the tab-separated file explained above. The file stem of the fasta files needs to match the last directory path string of the FTP path in column 3 of the input file (e.g. GCF_000002495.2_MG8)
output-filename
Path to the output file containing the hierarchical interleaved XOR filter index of the reference sequences and taxonomy information for the profiling step.
kmer-size
Size of k-length-substrings used for pseudo-mapping. When using syncmers for downsampling, the kmer-size has to be even-numbered because of using open canonical syncmers. The maximum supported k-mer size is 30.
syncmer-size
Size of the substrings used for selecting a k-mer for pseudo-mapping. The syncmer-size also has to be even-numbered because of the usage of open canonical syncmers. This number needs to be smaller than the k-mer size and the maximum supported size is 26.
use-syncmer
Switch that enables the usage of syncmers for downsampling of k-mers.
threads
Number of threads used for computing the hierarchical structure and building the HIXF index.
taxor-search - Queries a file of DNA sequences against an HIXF index
====================================================================
DESCRIPTION
Query sequences against the taxor HIXF index structure
OPTIONS
Basic options:
-h, --help
Prints the help page.
-hh, --advanced-help
Prints the help page including advanced options.
--version
Prints the version information.
--copyright
Prints the copyright/license information.
--export-help (std::string)
Export the help page information. Value must be one of [html, man].
Main options:
--index-file (std::string)
taxor index file containing HIXF index and reference sequence information
--query-file (std::string)
file containing sequences to query against the index Default: .
--output-file (std::string)
A file name for the resulting output. Default: .
--threads (unsigned 8 bit integer)
The number of threads to use. Default: 1. Value must be in range [1,32].
--percentage (double)
If set, this threshold is used instead of the k-mer/syncmer models. Default: -1. Value must be in range
[0,1].
--error-rate (double)
Expected error rate of reads that will be queried Default: 0.04. Value must be in range [0,1].
index-file
Path to the file containing the hierarchical interleaved XOR filter index of the reference sequences and taxonomy information for the profiling step.
query-file
Path to a fast(a/q) file containing sequenced reads of a sample, which shall be taxonomically classified. This file can be gzip compressed.
output-file
Path to the output file containing results of the classification step. This file is tab-separated with 10 columns per line.
- Column 1 : read identifier
- Column 2 : Assembly Accession ID of the matching reference
- Column 3 : Organism name of the matching reference
- Column 4 : Taxonomy ID of the matching reference
- Column 5 : Cumulative length of the matching reference sequence
- Column 6 : Sequence length of the queried read
- Column 7 : Overall number of k-mers (syncmers) generated from the queried read
- Column 8 : Number of k-mers (syncmers) of the query read that match with the reference sequence
- Column 9 : Taxonomy string of the matched reference
- Column 10 : Taxonomy ID string of the matched reference
#QUERY_NAME ACCESSION REFERENCE_NAME TAXID REF_LEN QUERY_LEN QHASH_COUNT QHASH_MATCH TAX_STR TAX_ID_STR
f3a88e6f-9873-445a-b0c7-2406f3a360fc|1613 GCF_022819245.1 Limosilactobacillus fermentum 1613 2016236 1139 98 56 k__Bacteria;p__Bacillota;c__Bacilli;o__Lactobacillales;f__Lactobacillaceae;g__Limosilactobacillus;s__Limosilactobacillus fermentum 2;1239;91061;186826;33958;2742598;1613
9d3a345f-3bf7-4304-977c-ee8b2b3c2b17|96241 GCF_000959025.1 Bacillus intestinalis 1963032 4031727 2589 224 104 k__Bacteria;p__Bacillota;c__Bacilli;o__Bacillales;f__Bacillaceae;g__Bacillus;s__Bacillus intestinalis 2;1239;91061;1385;186817;1386;1963032
9d3a345f-3bf7-4304-977c-ee8b2b3c2b17|96241 GCF_006094475.1 Bacillus spizizenii 96241 4045538 2589 224 104 k__Bacteria;p__Bacillota;c__Bacilli;o__Bacillales;f__Bacillaceae;g__Bacillus;s__Bacillus spizizenii 2;1239;91061;1385;186817;1386;96241
4e845c81-7397-4956-bdb2-5893b450aa3e|1613 GCF_022819245.1 Limosilactobacillus fermentum 1613 2016236 1790 158 38 k__Bacteria;p__Bacillota;c__Bacilli;o__Lactobacillales;f__Lactobacillaceae;g__Limosilactobacillus;s__Limosilactobacillus fermentum 2;1239;91061;186826;33958;2742598;1613
acc4abca-4daf-4af1-a3fc-edc189c37095|2807625 GCF_018622975.1 Staphylococcus sp. MZ9 2836367 2860948 3971 344 158 k__Bacteria;p__Bacillota;c__Bacilli;o__Bacillales;f__Staphylococcaceae;g__Staphylococcus;s__Staphylococcus sp. MZ9 2;1239;91061;1385;90964;1279;2836367
threads
Number of threads used for querying the sequences in the input file against the HIXF index.
percentage
Can be used to define the minimum percentage of k-mers/syncmers that need to match a reference. By default we use the k-mer model from Blanca et al. or empircally computed values for determining the thesholds for reporting a match.
error-rate
For more accurate classification of reads we are calculating the expected number of mutated k-mers for each read prefix based on the expected sequencing error rate. Than a confidence interval for the mutated k-mers is calculated as described by Blanca et al. and the minimum number of matching k-mers is calculated based on the upper bound of the confidence interval. The significance level of the confidence interval is set to 95% by default. When using synmcers, we are using emprically calculated minimum numbers of matching syncmers for a given error rate and k-mer length.
taxor-profile - Taxonomic profiling of a sample by giving read matching results of Taxor search
===============================================================================================
DESCRIPTION
Taxonomic profiling of the given read set
OPTIONS
Basic options:
-h, --help
Prints the help page.
-hh, --advanced-help
Prints the help page including advanced options.
--version
Prints the version information.
--copyright
Prints the copyright/license information.
--export-help (std::string)
Export the help page information. Value must be one of [html, man].
Main options:
--search-file (std::string)
taxor search file containing results of read querying against the HIXF index
--cami-report-file (std::string)
output file reporting genomic abundances in CAMI profiling format. This is the relative genome abundance in
terms of the genome copy number for the respective TAXID in the overall sample. Note that this is not
identical to the relative abundance in terms of assigned base pairs.
--seq-abundance-file (std::string)
output file reporting sequence abundance in CAMI profiling format (including unclassified reads). This is
the relative sequence abundance in terms of sequenced base pairs for the respective TAXID in the overall
sample. Note that this is not identical to the genomic abundance in terms of genome copy number for the
respective TAXID. Default: .
--binning-file (std::string)
output file reporting read to taxa assignments in CAMI binning format
--sample-id (std::string)
Identifier of the analyzed sample
--min-abundance (double)
Minimum abundance to report (default: 0.001) Default: 0.001. Value must be in range [0,1].
--em-steps (unsigned 16 bit integer)
The number of steps for the expectation maximization (EM) algorithm (default: 100). Default: 100. Value
must be in range [1,1000].
--threads (unsigned 8 bit integer)
The number of threads to use. Default: 1. Value must be in range [1,32].
SKiM classification options (only used with --classification-method skim):
--classification-method (std::string)
Read classification method: "em" (Taxor's own EM-based profiler, default) or "skim" (the
binomial-hypothesis-test classifier from Schneggenburger & Zola, "SKiM: accurately classifying metagenomic
ONT reads in limited memory"). Both can be run on the same --search-file (with different output files) to
compare them. Default: em. Value must be one of [em,skim].
--kmer-size (unsigned 8 bit integer)
k-mer size used when building the index (required for --classification-method skim; must match the value
used for taxor build/search). Default: 0. Value must be in range [1,64].
--skim-nfixed (unsigned 16 bit integer)
SKiM's fixed normalization sample size for the binomial test (default: 100, matching the paper). Default:
100. Value must be in range [1,100000].
--skim-cutoff-exponent (unsigned 16 bit integer)
SKiM's p-value cutoff, as a power of ten: cutoff = 10^-e (default: e=12, matching the paper). Default: 12.
Value must be in range [1,300].
search-file
Path to the output file of the search step containing results of the classification. This file is tab-separated with 10 columns per line as described above.
cami-report-file
Output file reporting genomic abundances in CAMI profiling format. This is the relative genome abundance or taxonomic abundance in terms of the genome copy number for the respective TAXID in the overall sample.
seq-abundance-file
Output file reporting sequence abundance in CAMI profiling format (including unclassified reads). This is the relative sequence abundance in terms of sequenced base pairs for the respective TAXID in the overall sample.
binning-file
Output file reporting read to taxon assignments in CAMI binning format.
sample-id
String that identifies the analyzed sample.
min-abundance
Taxa below this relative abundance are omitted from the report and sequence-abundance files.
em-steps
Maximum number of iterations for the EM algorithm (only used with --classification-method em); EM stops earlier if the log-likelihood converges.
threads
Number of threads used for taxonomic profiling.
By default, taxor profile assigns reads to taxa with its own EM-based
profiler: reference filtering (to suppress near-identical strain redundancy)
followed by iterative expectation-maximization over posterior probabilities.
Passing --classification-method skim switches to a second, independent
classifier implementing the read-classification algorithm from
Schneggenburger & Zola, "SKiM: accurately classifying metagenomic ONT reads
in limited memory".
Unlike the EM profiler, SKiM makes a single, direct statistical decision per
read with no iterative refinement and no cross-read preprocessing: for every
candidate reference it performs a binomial hypothesis test on the read's
k-mer/syncmer match count, classifying the read to the reference with the
lowest p-value (if it clears a cutoff), or leaving it unclassified otherwise.
Both methods read the same taxor search output and write the same three
CAMI-format output files, so they can be run back-to-back on the same
--search-file (with different output paths) to compare them directly:
taxor profile --search-file sample.search.txt --classification-method em \
--cami-report-file sample.em.report --seq-abundance-file sample.em.abundance \
--binning-file sample.em.binning --sample-id sample
taxor profile --search-file sample.search.txt --classification-method skim --kmer-size 20 \
--cami-report-file sample.skim.report --seq-abundance-file sample.skim.abundance \
--binning-file sample.skim.binning --sample-id sample
kmer-size
Required for --classification-method skim. Must match the --kmer-size used for taxor build/taxor search on
this index. SKiM's classification model needs the k-mer/syncmer universe size (derived from kmer-size) to compute
each reference's match probability; there is no cheap way to read this back out of the index file, so it must be
supplied explicitly.
skim-nfixed
SKiM normalizes each read's match count to a fixed number of trials (n_fixed) before the binomial test, so reads of
different lengths are compared on the same scale. The paper's default is 100; larger values give the test more
resolution for very long reads at the cost of a bigger per-reference precomputed probability table.
skim-cutoff-exponent
A read is classified to a reference only if that reference's binomial p-value is below 10^-e. The paper's default,
e=12, is intentionally strict (favoring precision over recall); lowering e classifies more reads at the risk of
more false positives.
Notes on the implementation: SKiM's model needs p_i = |K_i|/|U| (the fraction of the k-mer/syncmer universe
covered by a reference's indexed k-mers), but Taxor's search output only records each reference's sequence length
(REF_LEN), not its actual indexed k-mer/syncmer count. Since every reference in one index shares the same
kmer_size/syncmer_size/scaling, the syncmer-selection-density and scaling factors are common to both a
reference's true |K_i| and the universe size |U| and cancel out of the ratio, so Taxor approximates
p_i ≈ REF_LEN / (4^kmer_size / 2), using only REF_LEN (already available) and --kmer-size. This assumes roughly
uniform k-mer density and non-repetitive references; repeat-rich genomes get a systematically more conservative
(never more liberal) p-value as a result. Because SKiM is a direct per-read test with no cross-read signal, it
intentionally skips the EM profiler's reference-filtering/strain-deduplication preprocessing, to stay faithful to the
published method - this also means SKiM has no built-in defense against per-read assignment noise across
near-identical strain references the way the EM path does.
First download the reference sequences and taxonomy dump of the sequences from the NCBI using genome_updater.
genome_updater.sh \
-d "refseq"\
-g "archaea,bacteria,fungi,viral" \
-c "all" \
-l "complete genome,chromosome" \
-f "genomic.fna.gz" \
-o "refseq-abv" \
-t 12 \
-A "species:1" \
-m -a -p
Then, unpack the taxonomy dump and create a tab-separated-values file using the Linux command cut and taxonkit.
# cd to 2023-03-15_12-56-12
# taxdump
mkdir -p taxdump
tar -zxvf taxdump.tar.gz -C taxdump
cut -f 1,7,20 assembly_summary.txt \
| taxonkit lineage -i 2 -r -n -L --data-dir taxdump \
| taxonkit reformat -I 2 -P -t --data-dir taxdump \
| cut -f 1,2,3,4,6,7 > refseq_accessions_taxonomy.csv
Note (2024+ taxdump releases): NCBI has renamed the "superkingdom" rank to
"domain" for Bacteria/Archaea/Eukaryota, and Viruses now carry a different
rank ("acellular root") altogether. Since taxonkit reformat builds the
k__/p__/.../s__ lineage string by matching the rank literally named
"superkingdom", the resulting refseq_accessions_taxonomy.csv will have an
empty (or entirely missing) kingdom/domain entry for every organism. Taxor
itself won't fail on this - the kingdom-level rank will simply be absent from
the profiling output. To backfill it, run the bundled script, which
identifies the domain level via stable NCBI taxids (rank-name independent)
instead of relying on whatever NCBI currently calls the rank:
python3 scripts/fix_kingdom_rank.py \
--nodes taxdump/nodes.dmp \
--input refseq_accessions_taxonomy.csv \
--output refseq_accessions_taxonomy.fixed.csv
mv refseq_accessions_taxonomy.fixed.csv refseq_accessions_taxonomy.csv
Now we can build the hierarchical interleaved XOR filter (HIXF) index of the reference sequences and the NCBI taxonomy.
taxor build --input-file refseq_accessions_taxonomy.csv --input-sequence-dir refseq/2023-03-15_12-56-12/files \
--output-filename refseq-abfv-k22-s12.hixf --threads 6 --kmer-size 22 --syncmer-size 12 --use-syncmer
Then, we query the sample fastq file against the index allowing in this case a sequencing error rate of 15%.
taxor search --index-file refseq-abfv-k22-s12.hixf --query-file SAMPLE.fq.gz \
--output-file SAMPLE.search.txt --error-rate 0.15 --threads 6
Finally, the query result file is used as input for taxonomic profiling, which has three output files containing taxonomic abundances and sequence abundances in CAMI report format as well as a binning file with final read to reference assignments.
taxor profile --search-file SAMPLE.search.txt --cami-report-file SAMPLE.report \
--seq-abundance-file SAMPLE.abundance --binning-file SAMPLE.binning --sample-id SAMPLE