1. Jasmijn Baaijens
  2. savage



SAVAGE: Strain Aware VirAl GEnome assembly



SAVAGE is a computational tool for reconstructing individual haplotypes of intra-host virus strains (a viral quasispecies) without the need for a high quality reference genome. SAVAGE makes use of either FM-index based data structures or ad-hoc consensus reference sequence for constructing overlap graphs from patient sample data. In this overlap graph, nodes represent reads and/or contigs, while edges reflect that two reads/contigs, based on sound statistical considerations, represent identical haplotypic sequence. Following an iterative scheme, a new overlap assembly algorithm that is based on the enumeration of statistically well-calibrated groups of reads/contigs then efficiently reconstructs the individual haplotypes from this overlap graph.

For a more detailed description, please see our Genome Research paper or the preprint on BioRxiv.


A viral quasispecies, the ensemble of viral strains populating an infected person, can be highly diverse. For optimal assessment of virulence, pathogenesis and therapy selection, determining the haplotypes of the individual strains can play a key role. As many viruses are subject to high mutation and recombination rates, high- quality reference genomes are often not available at the time of a new disease outbreak. SAVAGE is able to reconstruct individual haplotypes of the viral quasispecies without the need for a high- quality reference genome.

Installation and dependencies

The easiest and recommended way to install SAVAGE is through Conda, using the Bioconda channel. Alternatively, SAVAGE can also be installed from source code. The installation procedure for either option is described below.

Please note that SAVAGE is built for linux-based systems only.

Installation via Conda

This is the easiest and recommended way to install SAVAGE. The Conda package manager allows you to install packages without needing any root privileges. First make sure to setup Bioconda by following the steps below:

  • Install the Miniconda Python2 distribution, see here.
  • Make sure that the miniconda2 bin directory is added to your PATH.
  • Setup the bioconda channel as well as the other channels bioconda depends on; it is important to add them in this order:
conda config --add channels r
conda config --add channels defaults
conda config --add channels conda-forge
conda config --add channels bioconda

More information about installing Bioconda can be found here.

Once Bioconda is properly setup, SAVAGE and all its dependencies can be installed through one simple command:

conda install savage

Installation from source

If installation via Conda is not an option for you, try installation from source. The SAVAGE program is a combination of C++ code (which has to be compiled) and Python scripts (version 2.6 or later, but not Python 3.x). The C++ part requires several boost libraries (boost::timer, boost::system, and boost::program_options) and it needs a compiler that supports OpenMP (such as g++).

For maximal clique enumeration, SAVAGE depends on the quick-cliques package which is already included.

As of version 0.4.0, SAVAGE uses the rust-overlaps package for computing suffix-prefix overlaps in de novo mode.

For reference-guided assembly, SAVAGE depends on the bwa mem aligner.

Stages b and c of the algorithm may require blastn and makeblastdb from the ncbi C++ Toolkit.

We provide a simple algorithm for relative frequency estimation of the assembled contigs. For improved frequency estimation and frequency-based filtering during assembly, we recommend using Kallisto. This is optional, for more information please see the frequency estimation and filtering sections below.

To summarize, please download and install:

Each of these tools can also be installed using Bioconda, a distribution of bioinformatics software realized as a channel for the versatile Conda package manager.

Once all dependencies are installed, download the latest version of SAVAGE, enter the repository and type make.


Currently, SAVAGE is designed for Illumina MiSeq sequencing reads that have been pre-processed by PEAR (for merging self-overlapping pairs). Thus, the input consists of (a combination of) single-end and/or paired-end reads. Important: SAVAGE expects paired-end reads stored in two fastq files with both sequences in forward orientation (instead of the more standard forward-reverse) because this how PEAR outputs non-merged reads. Please find a more detailed description of the SAVAGE input files below.

SAVAGE offers two modes: SAVAGE-ref and SAVAGE-de-novo. SAVAGE-ref is the more efficient option, while SAVAGE-de-novo gives (slightly) better results. Instructions for using these modes can be found below.

The most important parameter to set is the minimal overlap length M during overlap graph construction. For best results, please read the section on parameters carefully.


SAVAGE takes as input single-end and/or paired-end reads, which should be stored in separate fastq files (e.g. singles.fastq, paired1.fastq, paired2.fastq). The paired-end reads are assumed to be in two fastq files with both sequences in forward orientation (instead of the more standard forward-reverse) because this how PEAR outputs non-merged reads. These files can be specified using the following program options:

  • -s or --input_s Input fastq file containing single-end reads.
  • -p1 or --input_p1 Input fastq file containing /1 paired-end reads.
  • -p2 or --input_p2 Input fastq file containing /2 paired-end reads.
  • --ref (optional) Fasta file containing the reference genome to be used for reference-guided mode. Please make sure to enter the full path to this fasta file

The algorithm has a built-in error correction step, but it does assume the input data to be cleaned (i.e., adapter sequences should be removed and low quality bases should be trimmed). This can be done using cutadapt.

Assembly parameters

The two most important parameters when running SAVAGE are --split and --min_overlap_len, which should be chosen with great care. The first is always required when running SAVAGE, while the second will be estimated from the input data if not specified otherwise.

Before running SAVAGE, please read through the basic parameter settings described below.

  • --split (required) In case of (ultra-)deep sequencing data, exceeding a coverage of 1000x, we advise to split the data into patches of coverage between 500x and 1000x and run SAVAGE Stage a on each patch individually. After specifying the number of patches, SAVAGE takes care of the splitting and recombining. Choose the number of patches using such that 500 < read_coverage/patch_num < 1000.
  • -m, --min_overlap_len By default this parameter is set to 60% of the average length of the sequencing reads used as input for SAVAGE. The user can manually change this threshold using the -m or --min_overlap_len parameter. Increasing the minimal overlap length speeds up the algorithm and leads to a lower mismatch rate in the final contigs. It is recommended to set the minimal overlap length to be larger than the (expected) largest repetitive element in the target genomes. However, it also results in a lower fraction of the target genomes being reconstructed. We advise you to consider this trade-off when setting this parameter. We have been working with a minimal overlap length of 100-300bp, depending on the read length.
  • -t, --num_threads By default the number of threads is set to 1. Computationally-heavy parts of SAVAGE are multi-threaded and will profit from an increased value of -t. Especially in de novo mode it is highly recommended to use a larger number of threads.
  • --merge_contigs By default this is set to 0, meaning that in stage c, the final assembly step, we allow overlaps with a mismatch rate of 0% (i.e. exact overlaps). By increasing this threshold, e.g. to 0.01, virus strains which differ by less than 1% will be merged into one or more master strains, possibly leading to longer contigs and a less fragmented assembly (higher N50).
  • --sfo_mm This parameter is only relevant when running SAVAGE in de novo mode. It specifies the error rate allowed when computing approximate suffix-prefix overlaps. By default it is equal to 50, meaning that up to 2% mismatches is allowed in the overlaps. This accounts for 1% sequencing errors. Increasing this parameter will slow down the algorithm, while decreasing will lead to a possibly incomplete overlap graph.
  • --overlap_len_stage_c For Stage c of the algorithm, the final assembly step, it is possible to specify a different minimum overlap length using this option. By default this parameter is set to 100, but depending on the data it can pay off to decrease this parameter further.
  • --contig_len_stage_c By default, only contigs of at least 100 bp in length are considered for stage c assembly. The user can adjust this threshold by setting the contig_len_stage_c parameter. From the final stage c output, it is usually a good idea to consider only contigs of sufficient length, e.g. 500 bp.
  • --ignore_subreads By default this is set to False. When using the ignore_subreads flag, you choose not to use subread information from previous stages in the current stage(s). This will speed up the algorithm, but at the cost of less accurate abundance estimates.
  • --max_tip_len Maximum extension length for a sequence to be called a tip in the overlap graph. By default this parameter is set to the average length of the input sequences. If you want to disable tip removal, set --max_tip_len=0. In general this will lead to a more fragmented assembly and is therefore not recommended.
  • --no_filtering By default, SAVAGE contigs are filtered after stages b and c based on Kallisto frequency estimates: all zero-abundance contigs are removed from the assembly. To disable this filtering procedure, add the --no_filtering flag.


This mode constructs the overlap graph of the reads using the pairwise overlaps induced from a read-to-reference alignment. It takes as input a reference genome along with the read set, and uses bwa mem to compute alignments after splitting the data.

SAVAGE has proven to work not only on high-quality, well-curated reference genomes, but also on ad-hoc consensus genomes constructed from the read set itself using a de novo assembler (such as VICUNA). Therefore, even in ref-mode, SAVAGE can be used to compute very accurate de novo viral quasispecies assemblies.

To run SAVAGE-ref, complete the following steps:

  1. Download or assemble a reference genome reference.fasta.
  2. Create a directory my_directory where you want to store the results and enter this directory: cd my_directory.
  3. Now run SAVAGE with the option --ref /path/to/reference.fasta:
savage --ref /path/to/reference.fasta --split patch_num --min_overlap_len M --s singles.fastq --p1 paired1.fastq --p2 paired2.fastq


This mode does not rely on any reference sequence; instead, it computes all approximate suffix-prefix overlaps among the reads using the rust-overlaps package.

To run SAVAGE-de-novo, complete the following steps:

  1. Create a directory my_directory where you want to store the results and enter this directory: cd my_directory.
  2. Now run SAVAGE:
savage --split patch_num --min_overlap_len M --s singles.fastq --p1 paired1.fastq --p2 paired2.fastq

Algorithm stages

The algorithm proceeds in three stages: Stage a has the original reads as input and contigs. In this stage, SAVAGE performs error correction on the input reads. This has been optimized for a coverage of 500x-1000x, hence it is very important to set the --split parameter correctly. Stage b has these contigs as input and maximally extended contigs as output. Stage c attempts to merge maximized contigs into haplotypes. The parameter --merge_contigs allows you to specify the mismatch rate allowed within overlaps in this final stage. By increasing this threshold, e.g. to 0.01, virus strains which differ by less than 1% will be merged into one or more master strains*, possibly leading to longer contigs and a less fragmented assembly (higher N50).

By default, SAVAGE runs all three stages and the output of Stage c (contigs_stage_c.fasta) is considered as the viral quasispecies assembly.

Counting virus strains

When a sample contains strains with conserved regions larger than the read length, SAVAGE cannot connect variants across this region. In general, SAVAGE is very conservative: if it is not sure how to extend contigs, then it doesn't. As a consequence, SAVAGE may not be able to produce full-length haplotypes. For this reason, we provide an estimated (lower bound) on the number of strains in the sample. To apply this procedure, add the --count_strains flag to your SAVAGE command.

Note: to compute the estimated strain count a reference genome is required. The reference genome should be given through the parameter --ref. If you want to run SAVAGE in de novo mode it is required to do this separately. First run SAVAGE without specifying a reference genome and without the --count_strains flag. When the assembly is completed, run SAVAGE only for strain counting: on the command line, add flags --no_assembly and --count_strains and specify your reference genome with --ref /path/to/your/reference.fasta.

Frequency estimation

We provide a frequency estimation procedure which computes for every contig an estimate of its relative frequency. This procedure needs a fasta/fastq file of contigs and the corresponding subreads file (subreads.txt) as output by SAVAGE. This method works best when applied to Stage b or Stage c contigs, because those sequences have been maximally extended.

The frequency estimation can be applied as follows:

python freq_est.py --contigs contigs.fastq --subreads subreads.txt --min_len 1000

The parameter --min_len (or -m) can be set to any non-negative integer; only contigs of at least this length will be considered during frequency estimation.

For improved frequency estimation, we recommend to use Kallisto. This requires Kallisto to be installed on your machine. Kallisto uses the original paired-end sequencing reads (i.e., before running PEAR) from two separate files (forward.fastq, reverse.fastq). Run the following command to estimate relative frequencies in Kallisto mode:

freq_est --kallisto -f forward.fastq -r reverse.fastq --contigs contigs.fastq --min_len 1000

When using the Bioconda installation, please download freq_est.py manually from this repository.


The directory example/ contains a small working example, with both single- and paired-end reads contained in input_fas/ and an HIV-1 HXB2 reference strain (hiv-ref.fasta).

To run the example, cd into example/ and

  • for SAVAGE-ref execute:
savage --ref /path/to/example/hiv-ref.fasta -s input_fas/singles.fastq -p1 input_fas/paired1.fastq -p2 input_fas/paired2.fastq -m 200 --split 1
  • for SAVAGE-de-novo execute:
savage -s input_fas/singles.fastq -p1 input_fas/paired1.fastq -p2 input_fas/paired2.fastq -m 200 --split 1

High coverage data sets

In case of (ultra-)deep sequencing data, exceeding a coverage of 1000x, we advise to split the data into patches of coverage between 500x and 1000x. This is easily achieved by setting the --split parameter in your SAVAGE command. This will run SAVAGE Stage a on each patch individually, followed by a joint Stage b on the full set of contigs obtained from all patches, and subsequently also Stage c.

Make sure to choose the number of patches in which to split your data such that 500 < read_coverage/patch_num < 1000. Then set --split=patch_num and SAVAGE will take care of the rest.

Snakemake workflow

The earliest versions of SAVAGE provided a Snakefile to run experiments with Snakemake. However, this approach is no longer maintained; the complete SAVAGE workflow is now taken care of by savage.py. It is therefore recommended to run SAVAGE directly instead of using the Snakefile.

If you do wish to use the Snakefile, make sure that all SAVAGE parameters are specified in the config file savage_config.yaml. Please edit the config file and copy it to the directory where you wish to store all SAVAGE output, then run

snakemake -s /path/to/SAVAGE_ROOT/Snakefile

where /path/to/SAVAGE_ROOT/ specifies the path to the SAVAGE root directory. Note that the most recent options of SAVAGE are not supported by the Snakefile.

General remarks

SAVAGE expects as input single- and/or paired-end reads. For the paired-end reads, it assumes they are stored both on the same strand (hence resulting in F-F alignments) as output by PEAR, unless specified otherwise. If your reads are stored in forward-reverse orientations, make sure to use the option --revcomp when calling SAVAGE.


If you are using SAVAGE, please cite our paper: De novo viral quasispecies assembly using overlap graphs,
J. Baaijens, A. Zine El Aabidine, E. Rivals, and A. Schoenhuth,
Genome Res. 2017. 27: 835-848, doi:10.1101/gr.215038.116.


In case of any questions or issues, please contact Jasmijn Baaijens: <lastname> AT cwi DOT nl