freebayes, a haplotype-based variant detector
user manual and guide
FreeBayes is a Bayesian genetic variant detector designed to find small polymorphisms, specifically SNPs (single-nucleotide polymorphisms), indels (insertions and deletions), MNPs (multi-nucleotide polymorphisms), and complex events (composite insertion and substitution events) smaller than the length of a short-read sequencing alignment.
FreeBayes is haplotype-based, in the sense that it calls variants based on the literal sequences of reads aligned to a particular target, not their precise alignment. This model is a straightforward generalization of previous ones (e.g. PolyBayes, samtools, GATK) which detect or report variants based on alignments. This method avoids one of the core problems with alignment-based variant detection--- that identical sequences may have multiple possible alignments:
FreeBayes uses short-read alignments (BAM files with Phred+33 encoded quality scores, now standard) for any number of individuals from a population and a reference genome (in FASTA format) to determine the most-likely combination of genotypes for the population at each position in the reference. It reports positions which it finds putatively polymorphic in variant call file (VCF) format. It can also use an input set of variants (VCF) as a source of prior information, and a copy number variant map (BED) to define non-uniform ploidy variation across the samples under analysis.
A preprint Haplotype-based variant detection from short-read sequencing provides an overview of the statistical models used in FreeBayes. We ask that you cite this paper if you use FreeBayes in work that leads to publication.
Please use this citation format:
Garrison E, Marth G. Haplotype-based variant detection from short-read sequencing. arXiv preprint arXiv:1207.3907 [q-bio.GN] 2012
If possible, please also refer to the version number provided by freebayes when
it is run without arguments or with the
--help option. For example, you
should see something like this:
This provides both a point release number and a git commit id, which will ensure precise reproducibility of results.
To download FreeBayes, please use git to download the most recent development tree. Currently, the tree is hosted on github, and can be obtained via:
git clone --recursive git://github.com/ekg/freebayes.git
Note the use of --recursive. This is required in order to download all nested git submodules for external repositories.
After you've already done the above to clone the most recent development version, if you wish to compile a specific version of FreeBayes from, you can then do:
git checkout v0.9.18 && git submodule update --recursive
Resolving proxy issues with git
Depending on your local network configuration, you may have problems obtaining freebayes via git. If you see something like this you may be behind a proxy that blocks access to standard git:// port (9418).
$ git clone --recursive git://github.com/ekg/freebayes.git Cloning into 'freebayes'... fatal: Unable to look up github.com (port 9418) (Name or service not known)
Luckily, if you have access to https:// on port 443, then you can use this 'magic' command as a workaround to enable download of the submodules:
git config --global url.https://github.com/.insteadOf git://github.com/
FreeBayes requires g++ and the standard C and C++ development libraries. Additionally, cmake is required for building the BamTools API.
Will build the executable freebayes, as well as the utilities bamfiltertech and
bamleftalign. These executables can be found in the
bin/ directory in the
Users may wish to install to e.g. /usr/local/bin (default), which is accomplished via
sudo make install
In its simplest operation, freebayes requires only two inputs: a FASTA reference sequence, and a BAM-format alignment file sorted by reference position. For instance:
freebayes --fasta-reference h.sapiens.fasta NA20504.bam
... produce (on standard output) a VCF file on standard out describing all SNPs, INDELs, MNPs, and Complex events between the reference and the alignments in NA20504.bam. In order to produce correct output, the reference supplied must be the reference to which NA20504.bam was aligned.
Users may specify any number of BAM files on the command line. FreeBayes uses the BamTools API to open and parse these files in parallel, virtually merging them at runtime into one logical file with a merged header.
For a description of available command-line options and their defaults, run:
Call variants assuming a diploid sample:
freebayes -f ref.fa aln.bam >var.vcf
Require at least 5 supporting observations to consider a variant:
freebayes -f ref.fa -C 5 aln.bam >var.vcf
Use a different ploidy:
freebayes -f ref.fa -p 4 aln.bam >var.vcf
Assume a pooled sample with a known number of genome copies. Note that this
means that each sample identified in the BAM file is assumed to have 32 genome
copies. When running with highh --ploidy settings, it may be required to set
--use-best-n-alleles to a low number to limit memory usage.
freebayes -f ref.fa -p 32 --use-best-n-alleles 4 --pooled-discrete aln.bam >var.vcf
Generate frequency-based calls for all variants passing input thresholds:
freebayes -f ref.fa -F 0.01 -C 1 --pooled-continuous aln.bam >var.vcf
Use an input VCF (bgzipped + tabix indexed) to force calls at particular alleles:
freebayes -f ref.fa -@ in.vcf.gz aln.bam >var.vcf
Generate long haplotype calls over known variants:
freebayes -f ref.fa --haplotype-basis-alleles in.vcf.gz \ --haplotype-length 50 aln.bam
Naive variant calling: simply annotate observation counts of SNPs and indels:
freebayes -f ref.fa --haplotype-length 0 --min-alternate-count 1 \ --min-alternate-fraction 0 --pooled-continuous --report-monomorphic >var.vcf
Parallel operation (use 36 cores in this case cores):
freebayes-parallel <(fasta_generate_regions.py ref.fa.fai 100000) 36 \ -f ref.fa aln.bam >var.vcf
Note that any of the above examples can be made parallel by using the scripts/freebayes-parallel script. If you find freebayes to be slow, you should probably be running it in parallel using this script to run on a single host, or generating a series of scripts, one per region, and run them on a cluster.
Calling variants: from fastq to VCF
You've sequenced some samples. You have a reference genome or assembled set of contigs, and you'd like to determine reference-relative variants in your samples. You can use freebayes to detect the variants, following these steps:
- Align your reads to a suitable reference (e.g. with bwa or MOSAIK)
- Ensure your alignments have read groups attached so their sample may be identified by freebayes. Aligners allow you to do this, but you can also use bamaddrg to do so post-alignment.
- Sort the alignments (e.g. bamtools sort).
- Mark duplicates, for instance with samtools rmdup (if PCR was used in the preparation of your sequencing library)
- Run freebayes on all your alignment data simultaneously, generating a
VCF. The default settings should work for most use cases, but if your samples
are not diploid, set the
--ploidyand adjust the
- Filter the output e.g. using reported QUAL and/or depth (DP) or observation count (AO).
- Interpret your results.
- (possibly, Iterate the variant detection process in response to insight gained from your interpretation)
FreeBayes a standard VCF 4.1 outut stream. This format is designed for the probabilistic description of allelic variants within a population of samples, but it is equally suited to describing the probability of variation in a single sample.
Of primary interest to most users is the QUAL field, which estimates the probability that there is a polymorphism at the loci described by the record. In freebayes, this value can be understood as 1 - P(locus is homozygous given the data). It is recommended that users use this value to filter their results, rather than accepting anything output by freebayes as ground truth.
By default, records are output even if they have very low probability of
variation, in expectation that the VCF will be filtered using tools such as
vcflib, which is also included in the
vcflib/. For instance,
freebayes -f ref.fa aln.bam | vcffilter -f "QUAL > 20" >results.vcf
removes any sites with estimated probability of not being polymorphic less than phred 20 (aka 0.01), or probability of polymorphism > 0.99.
In simulation, the receiver-operator characteristic (ROC) tends to have a very sharp inflection between Q1 and Q30, depending on input data characteristics, and a filter setting in this range should provide decent performance. Users are encouraged to examine their output and both variants which are retained and those they filter out. Most problems tend to occur in low-depth areas, and so users may wish to remove these as well, which can also be done by filtering on the DP flag.
Calling variants in a population
FreeBayes is designed to be run on many individuals from the same population (e.g. many human individuals) simultaneously. The algorithm exploits a neutral model of allele diffusion to impute most-confident genotypings across the entire population. In practice, the discriminant power of the method will improve if you run multiple samples simultaneously. In other words, if your study has multiple individuals, you should run freebayes against them at the same time. This also ensures consistent reporting of information about evidence for all samples at any locus where any are apparently polymorphic.
To call variants in a population of samples, each alignment must have a read group identifier attached to it (RG tag), and the header of the BAM file in which it resides must map the RG tags to sample names (SM). Furthermore, read group IDs must be unique across all the files used in the analysis. One read group cannot map to multiple samples. The reason this is required is that freebayes operates on a virtually merged BAM stream provided by the BamTools API. If merging the files in your analysis using bamtools merge would generate a file in which multiple samples map to the same RG, the files are not suitable for use in population calling, and they must be modified.
Users may add RG tags to BAM files which were generated without this information by using (as mentioned in "Calling variants" above) bamaddrg. If you have many files corresponding to many individuals, add a unique read group and sample name to each, and then open them all simultaneously with freebayes. The VCF output will have one column per sample in the input.
If you find freebayes to be slow, or use large amounts of memory, consider the following options:
--use-best-n-alleles 4: this will reduce the number of alleles that are considered, which will decrease runtime at the cost of sensitivity to lower-frequency alleles at multiallelic loci. Calculating site qualities requires O(samples*genotypes) runtime, and the number of genotypes is exponential in ploidy and the number of alleles that are considered, so this is very important when working with high ploidy samples (and also
--pooled-discrete). By default, freebayes puts no limit on this.
--genotype-qualities: calculating genotype qualities requires O(samples*genotypes) memory.
Set higher input thresholds. Require that N reads in one sample support an allele in order to consider it:
--min-alternate-count N, or that the allele fraction in one sample is M:
--min-alternate-fraction M. This will filter noisy alleles. The defaults,
--min-alternate-count 2 --min-alternate-fraction 0.2, are most-suitable for diploid, moderate-to-high depth samples, and should be changed when working with different ploidy samples. Alternatively,
--min-alternate-qsumcan be used to set a specific quality sum, which may be more flexible than setting a hard count on the number of observations.
Observation filters and qualities
FreeBayes filters its input so as to ignore low-confidence alignments and
alleles which are only supported by low-quality sequencing observations (see
--min-base-quality). It also will only evaluate a
position if at least one read has mapping quality of
--min-supporting-mapping-quality and one allele has quality of at least
Reads with more than a fixed number of high-quality mismatches can be excluded
--read-mismatch-limit. This is meant as a workaround when
mapping quality estimates are not appropriately calibrated.
Reads marked as duplicates in the BAM file are ignored, but this can be
disabled for testing purposes by providing
does not mark duplicates on its own, you must use another process to do this.
As a guard against spurious variation caused by sequencing artifacts, positions
are skipped when no more than
non-clonal observations of an alternate are found in one sample. These default
to 2 and 0.2 respectively. The default setting of
--min-alternate-fraction 0.2 is suitable for diploid samples but should be changed for ploidy > 2.
Allele type exclusion
FreeBayes provides a few methods to ignore certain classes of allele, e.g.
--no-mnps. Users are strongly cautioned against using
these, because removing this information is very likely to reduce detection
power. To generate a report only including SNPs, use vcffilter post-call as
freebayes ... | vcffilter -f "TYPE = snp"
FreeBayes estimates observation quality using several simple heuristics based on manipulations of the phred-scaled base qualities:
- For single-base observations, mismatches and reference observations: the un-adjusted base quality provided in the BAM alignment record.
- For insertions: the mean quality of the bases inside of the putatively inserted sequence.
- For deletions: the mean quality of the bases flanking the putatively deleted sequence.
- For haplotypes: the mean quality of allele observations within the haplotype.
Effective base depth
By default, filters are left completely open.
--experimental-gls if you would like to integrate both base and mapping
quality are into the reported site quality (QUAL in the VCF) and
genotype quality (GQ, when supplying
--genotype-qualities). This integration
is driven by the "Effective Base Depth" metric first developed in
snpTools, which scales observation
quality by mapping quality. When
--experimental-gls is given, P(Obs|Genotype) ~
FreeBayes can read BAM from standard input
--stdin instead of directly from
files. This allows the application of any number of streaming BAM filters and
calibrators to its input.
bam_merger.sh | streaming_filter_or_process.sh | freebayes --stdin ...
This pattern allows the adjustment of alignments without rewriting BAM files, which could be expensive depending on context and available storage. A prime example of this would be graph-based realignment of reads to known variants as implemented in glia.
Using this pattern, you can filter out reads with certain criteria using bamtools filter without having to modify the input BAM file. You can also use the bamtools API to write your own custom filters in C++. An example filter is bamfiltertech [src/bamfiltertech.cpp](http://github.com/ekg/freebayes/blob/master/src/bamfilte rtech.cpp), which could be used to filter out technologies which have characteristic errors which may frustrate certain types of variant detection.
In principle, any gapped aligner which is sensitive to indels will produce satisfactory input for use by freebayes. Due to potential ambiguity, indels are not parsed when they overlap the beginning or end of alignment boundaries.
When calling indels, it is important to homogenize the positional distribution
of insertions and deletions in the input by using left realignment. This is
now done automatically by freebayes, but the behavior can be turned off via
--dont-left-align-indels flag. You probably don't want to do this.
Left realignment will place all indels in homopolymer and microsatellite repeats at the same position, provided that doing so does not introduce mismatches between the read and reference other than the indel. This method computationally inexpensive and handles the most common classes of alignment inconsistency.
As freebayes is haplotype-based, left-alignment is necessary only for the determination of candidate polymorphic loci. Once such loci are determined, haplotype observations are extracted from reads where:
- putative variants lie within
--haplotype-windowbases of each other (default 3bp),
- the reference sequence has repeats (e.g. microsatellites or STRs are called as one haplotype),
- the haplotype which is called has Shannon entropy less than
--min-repeat-entropy, which is off by default but can be set to ~1 for optimal genotyping of indels in lower-complexity sequence.
After a haplotype window is determined by greedily expanding the window across overlapping haplotype observations, all reads overlapping the window are used to establish data likelihoods, P(Observations|Genotype), for all haplotypes which have sufficient support to pass the input filters.
Partial observations are considered to support those haplotypes which they could match exactly. For expedience, only haplotypes which are contiguously observed by the reads are considered as putative alleles in this process. This differs from other haplotype-based methods, such as Platypus, which consider all possible haplotypes composed of observed component alleles (SNPs, indels) in a given region when generating likelihoods.
The primary adantages of this approach are conceptual simplicity and performance, and it is primarily limited in the case of short reads, an issue that is mitigated by increasing read lengths. Also, a hybrid approach must be used to call haplotypes from high-error rate long reads.
Re-genotyping known variants and calling long haplotypes
For longer reads with higher error rates, it is possible to generate long haplotypes in two passes over the data. For instance, if we had very long reads (e.g. >10kb) at moderate depth and high error rate (>5%) such as might be produced by PacBio, we could do something like:
freebayes -f ref.fa aln.bam | vcffilter -f "QUAL > 20" >vars.vcf
... thus generating candidate variants of suitable quality using the default detection window. We can then use these as "basis alleles" for the observation of haplotypes, considering all other putative variants supported by the alignment to be sequencing errors:
freebayes -f ref.fa --haplotype-window 500 \ --haplotype-basis-alleles vars.vcf aln.bam >haps.vcf
These steps should allow us to read long haplotypes directly from input data with high error rates.
The high error rate means that beyond a small window each read will contain a
completely different literal haplotype. To a point, this property improves our
signal to noise ratio and can effectively filter out sequencing errors at the
point of the input filters, but it also decreases the effective observation
depth will prevent the generation of any calls if a long
is combined with high a sequencing error rate.
Best practices and design philosophy
FreeBayes follows the patterns suggested by the Unix philosophy, which promotes the development of simple, modular systems that perform a single function, and can be combined into more complex systems using stream processing of common interchange formats.
FreeBayes incorporates a number of features in order to reduce the complexity of variant detection for researchers and developers:
- Indel realignment is accomplished internally using a read-independent method, and issues resulting from discordant alignments are dramatically reducedy through the direct detection of haplotypes.
- The need for base quality recalibration is avoided through the direct detection of haplotypes. Sequencing platform errors tend to cluster (e.g. at the ends of reads), and generate unique, non-repeating haplotypes at a given locus.
- Variant quality recalibration is avoided by incorporating a number of metrics, such as read placement bias and allele balance, directly into the Bayesian model. (Our upcoming publication will discuss this in more detail.)
A minimal pre-processing pipeline similar to that described in "Calling variants" should be sufficient for most uses. For more information, please refer to a recent post by Brad Chapman [on minimal BAM preprocessing methods](http://bcbio.wordpress.com/2013/10/21/updated-comparison-of-variant-det ection-methods-ensemble-freebayes-and-minimal-bam-preparation-pipelines/).
For a push-button solution to variant detection, from reads to variant calls, look no further than the gkno genome analysis platform.
FreeBayes is made by:
- Erik Garrison
- Thomas Sibley
- Dillon Lee
- Patrick Marks
- Noah Spies
- Joshua Randall
- Jeremy Anderson
If you would like to chat real-time about freebayes, join #freebayes on freenode. A gittr.im chat is also available.
Note that if you encounter issues with the development HEAD and you would like
a quick workaround for an issue that is likely to have been reintroduced
recently, you can use
git checkout to step back a few revisions.
git checkout [git-commit-id]
It will also help with debugging to know if a problem has arisen in recent commits!