Kraken taxonomic sequence classification system

Version 1.0

Operating Manual

Table of Contents

Introduction

Kraken is a taxonomic sequence classifier that assigns taxonomic labels to short DNA reads. It does this by examining the k-mers within a read and querying a database with those k-mers. This database contains a mapping of every k-mer in Kraken's genomic library to the lowest common ancestor (LCA) in a taxonomic tree of all genomes that contain that k-mer. The set of LCA taxa that correspond to the k-mers in a read are then analyzed to create a single taxonomic label for the read; this label can be any of the nodes in the taxonomic tree. Kraken is designed to be rapid, sensitive, and highly precise. Our tests on various real and simulated data have shown Kraken to have sensitivity slightly lower than Megablast with precision being slightly higher. On a set of simulated 100 bp reads, Kraken processed over 1.3 million reads per minute on a single core in normal operation, and over 4.1 million reads per minute in quick operation.

The latest released version of Kraken will be available at the Kraken website, and the latest updates to the Kraken source code are available at the Kraken GitHub repository.

If you use Kraken in your research, please cite the Kraken paper. Thank you!

System Requirements

Note: Users concerned about the disk or memory requirements should read the paragraph about MiniKraken, below.

Installation

To begin using Kraken, you will first need to install it, and then either download or create a database.

Kraken consists of two main scripts ("kraken" and "kraken-build"), along with several programs and smaller scripts. As part of the installation process, all scripts and programs are installed in the same directory. After installation, you can move the main scripts elsewhere, but moving the other scripts and programs requires editing the scripts and changing the "$KRAKEN_DIR" variables.

Once a directory is selected, you need to run the following command in the directory where you extracted the Kraken source:

./install_kraken.sh $KRAKEN_DIR

(Replace "$KRAKEN_DIR" above with the directory where you want to install Kraken's programs/directories.)

The install_kraken.sh script should compile all of Kraken's code and setup your Kraken data directory. Installation is successful if you see the message "Kraken installation complete."

Once installation is complete, you may want to copy the two main Kraken scripts into a directory found in your PATH variable (e.g., "$HOME/bin"):

cp $KRAKEN_DIR/bin/kraken $HOME/bin
cp $KRAKEN_DIR/bin/kraken-build $HOME/bin

After installation, you're ready to either create or download a database.

Kraken Databases

A Kraken database is a directory containing at least 4 files:

Other files may be present as part of the database build process.

In interacting with Kraken, you should not have to directly reference any of these files, but rather simply provide the name of the directory in which they are stored. Kraken allows both the use of a standard database as well as custom databases; these are described in the sections Standard Kraken Database and Custom Databases below, respectively.

Standard Kraken Database

NOTE: Building the standard Kraken database downloads and uses all complete bacterial, archeal, and viral genomes in Refseq at the time of the build. As of October 2017, this includes ~25,000 genomes, requiring 33GB of disk space. The build process will then require approximately 450GB of additional disk space. After building the standard database, usage of the database will require users to keep only the database.idx, database.kdb, and taxonomy/ files, which requires approximately 200GB of disk space. When running a sample against this database, users will need 175 GB of RAM. If you do not have this computational resources or require testing against this Refseq database of ~25,000 genomes, we recommend building a custom database with only the genomes needed for your application.

To create the standard Kraken database, you can use the following command:

kraken-build --standard --db $DBNAME

(Replace "$DBNAME" above with your preferred database name/location.)

This will download NCBI taxonomic information, as well as the complete genomes in RefSeq for the bacterial, archaeal, and viral domains. After downloading all this data, the build process begins; this is the most time-consuming step. If you have multiple processing cores, you can run this process with multiple threads, e.g.:

kraken-build --standard --threads 24 --db $DBNAME

Using 24 threads on a computer with 244 GB of RAM, the build process took approximately 5 hours (steps with an asterisk have some multi-threading enabled) in October 2017. Please note that the time required for building the database depends on the number of genomic sequences:

 
  24m50s  *Step 1 (create kmer set)
     n/a  Step 2 (reduce database, optional and skipped)
2h34m53s  *Step 3 (sort set)
     n/a  Step 4 (GI number to sequence ID map)
   0.17s  Step 5 (Sequence ID to taxon map)
 2h7m28s  *Step 6 (set LCA values)
--------
 5h7m11s  Total build time

Note that if any step (including the initial downloads) fails, the build process will abort. However, kraken-build will produce checkpoints throughout the installation process, and will restart the build at the last incomplete step if you attempt to run the same command again on a partially-built database.

After building the database, to remove any unnecessary files (including the library files no longer needed), run the following:

kraken-build --db $DBNAME --clean

To create a custom database, or to use a database from another source, see Custom Databases.

Notes for users with lower amounts of RAM:

  1. If you encounter problems with Jellyfish not being able to allocate enough memory on your system to run the build process, you can supply a smaller hash size to Jellyfish using kraken-build's --jellyfish-hash-size switch. Each space in the hash table uses approximately 6.9 bytes, so using "--jellyfish-hash-size 6400M" will use a hash table size of 6.4 billion spaces and require 44.3 GB of RAM.

  2. Kraken's build process will normally attempt to minimize disk writing by allocating large blocks of RAM and operating within them until data needs to be written to disk. However, this extra RAM usage may exceed your capacity. In such cases, you may want to use kraken-build's --work-on-disk switch. This will minimize the amount of RAM usage and cause Kraken's build programs to perform most operations off of disk files. This switch can also be useful for people building on a ramdisk or solid state drive. Please note that working off of disk files can be quite slow on some computers, causing builds to take several days if not weeks.

Custom Databases

We realize the standard database may not suit everyone's needs. Kraken also allows creation of customized databases.

To build a custom database:

  1. Install a taxonomy. Usually, you will just use the NCBI taxonomy, which you can easily download using:

    kraken-build --download-taxonomy --db $DBNAME

    This will download the sequence ID to taxon map, as well as the taxonomic name and tree information from NCBI. These files can be found in $DBNAME/taxonomy/ . If you need to modify the taxonomy, edits can be made to the names.dmp and nodes.dmp files in this directory; the gi_taxid_nucl.dmp file will also need to be updated appropriately.

  2. Install a genomic library. Four sets of standard genomes are made easily available through kraken-build:

    To download and install any one of these, use the --download-library switch, e.g.:

    kraken-build --download-library bacteria --db $DBNAME

    Other genomes in a FASTA/multi-FASTA file can also be added:

  3. Once your library is finalized, you need to build the database. Depending on your size requirements, you may want to adjust the k-mer and/or minimizer lengths from the defaults. Except for some small bookkeeping fields, a Kraken database will use sD + 8(4M) bytes, where s is the number of bytes used to store the k-mer/taxon pair (usually 12, but lower for smaller k-mers), D is the number of distinct k-mers in your library and M is the length (in bp) of the minimizers. Although D does increase as k increases, it is impossible to know exactly how many distinct k-mers will exist in a library for a given k without actually performing the count. By default, k = 31 and M = 15.

    The minimizers serve to keep k-mers that are adjacent in query sequences close to each other in the database, which allows Kraken to exploit the CPU cache. Changing the value of M can significantly affect the speed of Kraken, and neither increasing or decreasing M will guarantee faster or slower speed.

    To build the database, you'll use the --build switch:

    kraken-build --build --db $DBNAME

    As noted above, you may want to also use any of --threads, --kmer-len, or --minimizer-len to adjust the database build time and/or final size.

  4. Shrinking the database: The "--shrink" task allows you to take an existing Kraken database and create a smaller MiniKraken database from it. The use of this option removes all but a specified number of k-mer/taxon pairs to create a new, smaller database. For example:

    kraken-build --shrink 10000 --db $DBNAME --new-db minikraken

    This will create a new database named minikraken that contains 10000 k-mers selected from across the original database ($DBNAME).

    The --shrink task is only meant to be run on a completed database. However, if you know before you create a database that you will only be able to use a certain amount of memory, you can use the --max-db-size switch for the --build task to provide a maximum size (in GB) for the database. This allows you to create a MiniKraken database without having to create a full Kraken database first.

A full list of options for kraken-build can be obtained using kraken-build --help.

After building a database, if you want to reduce the disk usage of the database you can use kraken-build's --clean switch to remove all intermediate files from the database directory.

Classification

To classify a set of sequences (reads), use the kraken command:

kraken --db $DBNAME seqs.fa

Output will be sent to standard output by default. The files containing the sequences to be classified should be specified on the command line. Sequences can also be provided through standard input using the special filename /dev/fd/0.

Note that to obtain optimum speeds, Kraken's database should be loaded into RAM first. This can be done through use of a ramdisk, if you have superuser permissions. Failing that, you can use the --preload switch to kraken, e.g.:

kraken --preload --db $DBNAME seqs.fa

The database files will be loaded before classification using this switch. See Memory Usage and Efficiency for more information.

The kraken program allows several different options:

To get a full list of options, use kraken --help.

Output Format

Each sequence classified by Kraken results in a single line of output. Output lines contain five tab-delimited fields; from left to right, they are:

  1. "C"/"U": one letter code indicating that the sequence was either classified or unclassified.
  2. The sequence ID, obtained from the FASTA/FASTQ header.
  3. The taxonomy ID Kraken used to label the sequence; this is 0 if the sequence is unclassified.
  4. The length of the sequence in bp.
  5. A space-delimited list indicating the LCA mapping of each k-mer in the sequence. For example, "562:13 561:4 A:31 0:1 562:3" would indicate that:

For users who want the full taxonomic name associated with each input sequence, we provide a script named kraken-translate that produces two different output formats for classified sequences. The script operates on the output of kraken, like so:

kraken --db $DBNAME sequences.fa > sequences.kraken
kraken-translate --db $DBNAME sequences.kraken > sequences.labels

(The same database used to run kraken should be used to translate the output; see Kraken Environment Variables below for ways to reduce redundancy on the command line.)

The file sequences.labels generated by the above example is a text file with two tab-delimited columns, and one line for each classified sequence in sequences.fa; unclassified sequences are not reported by kraken-translate. The first column of kraken-translate's output are the sequence IDs of the classified sequences, and the second column contains the taxonomy of the sequence. For example, an output line from kraken of:

C     SEQ1    562     36      562:6

Would result in a corresponding output line from kraken-translate of:

SEQ1  root;cellular organisms;Bacteria;Proteobacteria;Gammaproteobacteria;Enterobacteriales;Enterobacteriaceae;Escherichia;Escherichia coli

Alternatively, kraken-translate accepts the option --mpa-format which will report only levels of the taxonomy with standard rank assignments (superkingdom, kingdom, phylum, class, order, family, genus, species), and uses pipes to delimit the various levels of the taxonomy. For example, kraken-translate --mpa-format --db $DBNAME with the above example output from kraken would result in the following line of output:

SEQ1  d__Bacteria|p__Proteobacteria|c__Gammaproteobacteria|o__Enterobacteriales|f__Enterobacteriaceae|g__Escherichia|s__Escherichia_coli

Taxonomy assignments above the superkingdom (d__) rank are represented as just "root" when using the --mpa-report option with kraken-translate.

Paired Reads

Kraken will classify paired reads when the user specifies the --paired option by first concatenating the reads using | before classifying the combined reads against the Kraken database.

A number of other options are included in Kraken v1.0 that simplifies analysis of the paired reads. The following describes these options and lists the possible combinations of these options and their behavior when applied. Note that all options require that the --paired option is specified and that two input FASTA/FASTQ files are provided.

Memory Usage and Efficiency

Kraken's execution requires many random accesses to a very large file. To obtain maximal speed, these accesses need to be made as quickly as possible. This means that the database must be in physical memory during execution. Although we provide the --preload option to Kraken for users who cannot use a ramdisk, the ramdisk is likely the simplest option, and is well-suited for installations on computers where Kraken is to be run a majority of the time. In addition, using a ramdisk allows the initial start-up of Kraken to be accomplished much more quickly. If a ramdisk is used, the --preload switch should not be used.

We also note that in some cases, --preload may not be needed (or even advisable). If you know that your database is already in memory (for example, if it has been recently read or unzipped, then it should be in your operating system cache, which resides in physical memory), then there is no need to perform this step. We have noticed that in low-memory (~8 GB) situations, preloading a MiniKraken DB is actually much slower than simply using cat minikraken/database.* > /dev/null. The selection of the best way to get the database into memory is dependent on several factors, including your total amount of RAM, operating system, and current free memory. For this reason, you may need to experiment with your own setup to find a good solution for you.

To create a ramdisk, you will need to have superuser (root) permission. As root, you can use the following commands to create a ramdisk:

mkdir /ramdisk
mount -t ramfs none /ramdisk

Optionally, you may have a trusted user who you want to be able to copy databases into this directory. In that case, you'll need to make that user the owner of the directory via chown.

To put the database on the ramdisk, simply copy the database directory to the ramdisk directory:

cp -a $DBNAME /ramdisk

And then you can use it with Kraken by specifying the database copy on the ramdisk, e.g.:

kraken --db /ramdisk/$DBNAME seqs.fa

Note that anything copied into a ramdisk will be deleted if the ramdisk is unmounted or the computer is restarted, so make sure that you have a copy of the database on a hard disk (or other non-volatile storage).

Note that when using the --paired option, Kraken will not (by default) make any attempt to ensure that the two files you specify are indeed matching sets of paired-end reads. To verify that the names of each read do indeed match, you can use the --check-names option in combination with the --paired option.

Sample Reports

To get an idea as to Kraken's results across an entire sample, we provide the kraken-report script. It is used like this:

kraken-report --db $DBNAME kraken.output

Note that the database used must be the same as the one used to generate the output file, or the report script may encounter problems. Output is sent to standard output.

The output of kraken-report is tab-delimited, with one line per taxon. The fields of the output, from left-to-right, are as follows:

  1. Percentage of reads covered by the clade rooted at this taxon
  2. Number of reads covered by the clade rooted at this taxon
  3. Number of reads assigned directly to this taxon
  4. A rank code, indicating (U)nclassified, (D)omain, (K)ingdom, (P)hylum, (C)lass, (O)rder, (F)amily, (G)enus, or (S)pecies. All other ranks are simply '-'.
  5. NCBI taxonomy ID
  6. indented scientific name

The scientific names are indented using spaces, according to the tree structure specified by the taxonomy.

By default, taxa with no reads assigned to (or under) them will not have any output produced. However, if you wish to have all taxa displayed, you can use the --show-zeros switch to do so. This can be useful if you are looking to do further downstream analysis of the reports, and want to compare samples. Sorting by the taxonomy ID (using sort -nf5) can provide a consistent line ordering between reports.

In addition, we also provide the program kraken-mpa-report; this program provides output in a format similar to MetaPhlAn's tab-delimited output. For kraken-mpa-report, multiple Kraken output files can be specified on the command line and each will be treated as a separate sample. For each taxon at the standard ranks (from domain to species), the count of reads in each sample assigned to any node in the clade rooted at that taxon is displayed. kraken-mpa-report is run in the same manner as kraken-report, and its output is also sent to standard output.

Confidence Scoring

At present, we have not yet developed a confidence score with a solid probabilistic interpretation for Kraken. However, we have developed a simple scoring scheme that has yielded good results for us, and we've made that available in the kraken-filter script. The approach we use allows a user to specify a threshold score in the [0,1] interval; the kraken-filter script then will adjust labels up the tree until the label's score (described below) meets or exceeds that threshold. If a label at the root of the taxonomic tree would not have a score exceeding the threshold, the sequence is called unclassified by kraken-filter.

A sequence label's score is a fraction C/Q, where C is the number of k-mers mapped to LCA values in the clade rooted at the label, and Q is the number of k-mers in the sequence that lack an ambiguous nucleotide (i.e., they were queried against the database). Consider the example of the LCA mappings in Kraken's output given earlier:

"562:13 561:4 A:31 0:1 562:3" would indicate that:

In this case, ID #561 is the parent node of #562. Here, a label of #562 for this sequence would have a score of C/Q = (13+3)/(13+4+1+3) = 16/21. A label of #561 would have a score of C/Q = (13+4+3)/(13+4+1+3) = 20/21. If a user specified a threshold over 16/21, kraken-filter would adjust the original label from #562 to #561; if the threshold was greater than 20/21, the sequence would become unclassified.

kraken-filter is used like this:

kraken-filter --db $DBNAME [--threshold NUM] kraken.output

If not specified, the threshold will be 0. kraken-filter's output is similar to kraken's, but a new field between the length and LCA mapping list is present, indicating the new label's score (or the root label's score if the sequence has become unclassified).

To give some guidance toward selecting an appropriate threshold, we show here the results of different thresholds on the MiSeq metagenome from the Kraken paper (see the paper for more details; note that the database used here is more recent than that used in the paper). Precision, sensitivity, and F-score are measured at the genus rank:

Thres Prec Sens F-score
0 95.43 77.32 85.43
0.05 97.28 76.31 85.53
0.10 98.25 75.13 85.15
0.15 98.81 73.87 84.54
0.20 99.13 72.82 83.96
0.25 99.38 71.74 83.33
0.30 99.55 70.75 82.71
0.35 99.61 69.53 81.90
0.40 99.66 68.35 81.09
0.45 99.70 66.93 80.09
0.50 99.71 65.49 79.06

As can be seen, with no threshold (i.e., Kraken's original labels), Kraken's precision is fairly high, but it does increase with the threshold. Diminishing returns apply, however, and there is a loss in sensitivity that must be taken into account when deciding on the threshold to use for your own project.

Kraken Environment Variables

The Kraken programs (with the exception of kraken-build) support the use of some environment variables to help in reducing command line lengths:

Upgrading Databases to v0.10+

The minimizer ordering in Kraken versions prior to v0.10.0-beta was a simple lexicographical ordering that provided a suboptimal distribution of k-mers within the bins. Ideally, the bin sizes would be uniform, but simple lexicographical ordering creates a bias toward low-complexity minimizers. To resolve this, the ordering is now "scrambled" by XORing all minimizers with a predefined constant to toggle half of each minimizer's bits before sorting. The more evenly distributed bins provide better caching performance, but databases created in this way are not compatible with earlier versions of Kraken. Kraken versions from v0.10.0-beta up to (but not including) v1.0 will support the use of the older databases, but we nonetheless recommend one of the two following options:

  1. Build a new database. This is the preferred option, as a newly-created database will have the latest genomes and NCBI taxonomy information.

  2. Re-sort an existing database. If you have a custom database, you may want to simply reformat the database to provide you with Kraken's increased speed. To do so, you'll need to do the following:

    kraken-build --upgrade --db $DBNAME

    (Note: the --threads switch is both valid and encouraged with this operation.)

    This command will not delete your existing $DBNAME/database.* files, but will simply rename them. If you're satisfied with the new database's performance, then you can use kraken-build's --clean option to remove the old files and save space.

    Sorting the database is step 3 of the build process, so you should expect a database upgrade to take about as long as step 3 took when building the original database.

Note that the rest of Kraken v0.10.0-beta's speed improvements are available without upgrading or changing your database.

Upgrading Databases to v1.0

Upgrading to the Kraken version 1.0 does not require rebuilding of any existing Kraken databases. The main updates for this version are within the building process itself. Due to the phasing out of NCBI GI numbers, Kraken version 1.0 does not rely on GI numbers and rather uses the sequence ID to taxon ID maps provided in the NCBI taxonomy. The new version of Kraken uses these in the building of the database but the final database files have not changed. Other changes include changes in the rsync downloads of Refseq databases and in updated runtimes.