Skip to content


Repository files navigation


Latest GitHub release Build status for gcc/clang Published in Genome Research

Consensus module for raw de novo DNA assembly of long uncorrected reads.


Racon is intended as a standalone consensus module to correct raw contigs generated by rapid assembly methods which do not include a consensus step. The goal of Racon is to generate genomic consensus which is of similar or better quality compared to the output generated by assembly methods which employ both error correction and consensus steps, while providing a speedup of several times compared to those methods. It supports data produced by both Pacific Biosciences and Oxford Nanopore Technologies.

Racon can be used as a polishing tool after the assembly with either short accurate data or data produced by third generation of sequencing. The type of data inputted is automatically detected. Although, Racon expects single-end short reads, while paired-end reads should be renamed with unique names up to the first whitespace and joined into a single file before mapping (which can be done with misc/

Racon takes as input only three files: contigs in FASTA/FASTQ format, reads in FASTA/FASTQ format and overlaps/alignments between the reads and the contigs in MHAP/PAF/SAM format. Output is a set of polished contigs in FASTA format printed to stdout. All input files can be compressed with gzip (which will have impact on parsing time).

Racon can also be used as a read error-correction tool. In this scenario, the MHAP/PAF/SAM file needs to contain pairwise overlaps between reads including dual overlaps.

A wrapper script is also available to enable easier usage to the end-user for large datasets. It has the same interface as racon but adds two additional features from the outside. Sequences can be subsampled to decrease the total execution time (accuracy might be lower) while target sequences can be split into smaller chunks and run sequentially to decrease memory consumption. Both features can be run at the same time as well.


  1. gcc 4.8+ or clang 3.4+
  2. cmake 3.2+
  3. zlib

CUDA Support

  1. gcc 5.0+
  2. cmake 3.10+
  3. CUDA 9.0+


To install Racon run the following commands:

git clone && cd racon && mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release .. && make

After successful installation, an executable named racon will appear in build/bin (alongside unit tests racon_test).

Optionally, you can run sudo make install to install racon executable to your machine.

To build the wrapper script add -Dracon_build_wrapper=ON while running cmake. After installation, an executable named racon_wrapper (python script) will be created in build/bin.

CUDA Support

Racon makes use of NVIDIA's GenomeWorks SDK for CUDA accelerated polishing and alignment.

To build racon with CUDA support, add -Dracon_enable_cuda=ON while running cmake. If CUDA support is unavailable, the cmake step will error out. Note that the CUDA support flag does not produce a new binary target. Instead it augments the existing racon binary itself.

cd build
cmake -DCMAKE_BUILD_TYPE=Release -Dracon_enable_cuda=ON ..

Note: Short read polishing with CUDA is still in development!


To generate a Debian package for racon, run the following command from the build folder -

make package


Usage of racon is as following:

racon [options ...] <sequences> <overlaps> <target sequences>

    # default output is stdout
        input file in FASTA/FASTQ format (can be compressed with gzip)
        containing sequences used for correction
        input file in MHAP/PAF/SAM format (can be compressed with gzip)
        containing overlaps between sequences and target sequences
    <target sequences>
        input file in FASTA/FASTQ format (can be compressed with gzip)
        containing sequences which will be corrected

    -u, --include-unpolished
        output unpolished target sequences
    -f, --fragment-correction
        perform fragment correction instead of contig polishing (overlaps
        file should contain dual/self overlaps!)
    -w, --window-length <int>
        default: 500
        size of window on which POA is performed
    -q, --quality-threshold <float>
        default: 10.0
        threshold for average base quality of windows used in POA
    -e, --error-threshold <float>
        default: 0.3
        maximum allowed error rate used for filtering overlaps
        disables consensus trimming at window ends
    -m, --match <int>
        default: 3
        score for matching bases
    -x, --mismatch <int>
        default: -5
        score for mismatching bases
    -g, --gap <int>
        default: -4
        gap penalty (must be negative)
    -t, --threads <int>
        default: 1
        number of threads
        prints the version number
    -h, --help
        prints the usage

only available when built with CUDA:
    -c, --cudapoa-batches <int>
        default: 0
        number of batches for CUDA accelerated polishing per GPU
    -b, --cuda-banded-alignment
        use banding approximation for polishing on GPU. Only applicable when -c is used.
    --cudaaligner-batches <int>
        default: 0
        number of batches for CUDA accelerated alignment per GPU
    --cudaaligner-band-width <int>
        default: 0
        Band width for cuda alignment. Must be >= 0. Non-zero allows user defined
        band width, whereas 0 implies auto band width determination.

racon_test is run without any parameters.

Usage of racon_wrapper equals the one of racon with two additional parameters:

    --split <int>
        split target sequences into chunks of desired size in bytes
    --subsample <int> <int>
        subsample sequences to desired coverage (2nd argument) given the
        reference length (1st argument)

Contact information

For additional information, help and bug reports please send an email to one of the following:,,,


This work has been supported in part by Croatian Science Foundation under the project UIP-11-2013-7353. IS is supported in part by the Croatian Academy of Sciences and Arts under the project "Methods for alignment and assembly of DNA sequences using nanopore sequencing data". NN is supported by funding from A*STAR, Singapore.