SGA - String Graph Assembler

SGA is a de novo assembler for DNA sequence reads. It is based on Gene Myers' string graph
formulation of assembly and uses the FM-index/Burrows-Wheeler transform to efficiently
find overlaps between sequence reads. The core algorithms are described in this paper:

Compiling SGA

SGA dependencies:
    -google sparse hash library (
    -the bamtools library (
    -zlib (
    -(optional but suggested) the jemalloc memory allocator (

Additionally, the pipeline python scripts use the following modules. These
are not required to build SGA but must be available if you want to use the relevant 
python helper scripts:
    -pysam (
    -ruffus (

If you cloned the repository from github, run from the src directory 
to generate the configure file:


If bamtools and the sparsehash have been installed in standard locations (like /usr/local) you
can run configure without any parameters then run make:


If bamtools or the sparsehash are installed elsewhere, you can specify their locations as follows:

./configure --with-sparsehash=/home/jsimpson/ --with-bamtools=/home/jsimpson/software/bamtools

These directories should be the root of the install (in other words, the directories have
include/ and lib/ as subdirectories contained the header files and libraries, respectively).

The program uses pthreads to parallelize most steps of the assembly. The use of a concurrent
memory allocator like jemalloc can improve running time. If you would like to enable
the jemalloc memory allocator, specify the path as follows:

./configure --with-jemalloc=/home/jsimpson/jemalloc/lib/

The Hoard and tcmalloc allocators are also supported and can be enabled 
using --with-hoard/--with-tcmalloc if desired.

Installing SGA

Running make install will install sga into /usr/local/bin/ by default. To specify the install
location use the --prefix option to configure:

./configure --prefix=/home/jsimpson/ && make && make install

This command will copy sga to /home/jsimpson/bin/sga

Running SGA

SGA consists of a number of subprograms, together which form the assembly pipeline. The subprograms
can also be used to perform other interesting tasks, like read error correction or removing PCR duplicates.
Each program and subprogram will print a brief description and its usage instructions if the --help flag is used.

To get a listing of all subprograms, run sga --help.

Examples of an SGA assembly are provided in the src/examples directory. It is suggested to look at
these examples to become familar with the flow of data through the program.

The major subprograms are:

* sga preprocess 

Prepare reads for assembly. It can perform optional quality filtering/trimming. By default
it will discard reads that have uncalled bases ('N' or '.'). If you wish to keep these reads, use the --permuteAmbiguous
flag which will randomly change any uncalled bases to one of [ACGT]. It is mandatory to run this command 
on real data. 

If your reads are paired, the --pe-mode option should be specified. The paired reads can be input in two 
files (sga preprocess READS1 READS2) where the first read in READS1 is paired with the first read on READS2 
and so on. Alternatively, they can be specified in a single file where the two reads are expected to appear 
in consecutive records. By default, output is written to stdout.

* sga index -a ropebwt READS

Build the FM-index for READS, which is a fasta or fastq file. 
This program is threaded (-t N).

* sga correct READS

Perform error correction on READS file. Overlap and kmer-based correction algorithms
are implemented. By default, a k-mer based correction is performed. 

Many options exist for this program, see --help. 

* sga filter READS

Remove exact-match duplicated sequences and reads with low-frequency k-mers.
This program automatically regenerates the FM-index without the removed reads.

* sga overlap -m N READS

Find overlaps between reads to construct the string graph. The -m parameter specifies
the minimum length of the overlaps to find. By default only non-transitive (irreducible) edges are output and edges
between identical sequences. If all overlaps between reads are desired, the --exhaustive option can be specified.
This program is threaded. The output file is READS.asqg.gz by default.

* sga assemble READS.asqg.gz

Assemble takes the output of the overlap step and constructs contigs. The output is in contigs.fa by default. Options
exist for cleaning the graph before assembly which will substantially increase assembly continuity. 
See the --cut-terminal, --bubble, --resolve-small options.

Data quality issues

Sequence assembly requires high quality data. It is worth assessing the quality of your reads
using tools like FastQC ( to help guide the choice
of assembly parameters. Low-quality data should be filtered or trimmed.

Very highly-represented sequences (>1000X) can cause problems for SGA. This can happen when sequencing a small genome
or when mitochondria or other contamination is present in the sequencing run. In these cases, it is worth considering
pre-filtering the data or running an initial 'rmdup' step before error correction.


The first SGA code check-in was August, 2009. The algorithms for directly constructing the string graph from
the FM-index were developed and implemented in the fall of 2009. The initial public release was October 2010.


SGA is licensed under GPLv3. See the COPYING file in the src directory.

Third party code

SGA uses Bentley and Sedgwick's multikey quicksort code that can be found here:
It also uses zlib, the google sparse hash and gzstream by Deepak Bandyopadhyay and Lutz Kettner (see Thirdparty/README).
SGA also uses the stdaln dynamic programming functions written by Heng Li. The ropebwt implementation is by Heng Li

SGA also uses the rapidjson library by Milo Yip (

This Software includes an implementation of the algorithm published by Bauer et al. 
[Markus J. Bauer, Anthony J. Cox and Giovanna Rosone: Lightweight BWT Construction 
for Very Large String Collections. Lecture Notes in Computer Science 6661 
(Proceedings of the 22nd Annual Conference on Combinatorial Pattern Matching), 
2011, pp.219-231]. Intellectual property rights in the algorithm are owned and 
protected by Illumina Cambridge Ltd. and/or its affiliates (“Illumina”). 
Illumina is not a contributor or licensor under this GPL license and 
retains all rights, title and interest in the algorithm and intellectual 
property associated therewith. No rights or licenses to Illumina intellectual 
property, including patents, are granted to you under this license.


Written by Jared Simpson.
The algorithms were developed by Jared Simpson and Richard Durbin.