Peter Krusche pkrusche@illumina.com
This is a set of programs based on htslib to benchmark variant calls against gold standard truth datasets.
To compare a VCF against a gold standard dataset, use the following commmand line to perform genotype-level haplotype comparison.
hap.py truth.vcf query.vcf -f confident.bed -o output_prefix -r reference.faWe also have a script to perform comparisons only based on chromosome, position, and allele identity. This comparison will not resolve haplotypes and only verify that the same alleles were observed at the same positions (e.g. for comparison of somatic callsets).
som.py truth.vcf query.vcf -f confident.bed -o output_prefix -r reference.faMore information can be found below in the usage section.
A major challenge when comparing VCF files for diploid samples is the handling of complex variant representations. In a VCF file, we describe two haplotype sequences by means of REF-ALT pairs and genotypes. These variant calls do not uniquely represent the haplotype sequences: since alignments are always not unique even when using a fixed set of gap and substitution scores, different variant calling methods may produce different variant representations. While some of these representational differences can be handled using pre-processing of VCF files (e.g. variant trimming and left-shifting), others cannot be fixed easily.
In addition to comparing VCF records individually, we produce a graph-based representation of the VCF alleles, create all possible haplotype sequences, and compare these by alignment / exact matching. Here is an example where this is needed:
Variant representation 1 (shown in purple in the image below):
CHROM POS REF ALT GT
chrQ 10 G GTGTGTGCATGCT 0/1
Variant representation 2 (shown in green in the image below):
CHROM POS REF ALT GT
chrQ 16 G GCATGCT 0/1
chrQ 19 T TGTGTG 0/1
Both representations in this example are able to produce the same alt sequences, but we are not able to match them up with standard VCF tools. In particular, we can see from this example that the second representation actually may allow us to create two different sets of alt sequences if they are part of unphased heterozygous variant calls. When we don't know the phasing of our variants, the insertions could have occurred on different haplotypes when using representation 2.
With this tool, we can produce all haplotypes sequences by enumerating paths through a reference graph. By finding the paths / alt alleles that are consistent between two VCFs files we can produce accurate benchmarking numbers for comparing a VCF to a gold standard truth set. See doc/spec.md for more information.
An alternative method to compare complex variant calls is implemented in RTG vcfeval. It is possible to use vcfeval with hap.py, and to use hap.py only for pre-processing, stratification and counting.
The comparison method in vcfeval is more sophisticated than ours and can resolve some corner cases more accurately. For whole-genome comparisons, the difference between the two benchmarking methods is small, but when focusing on difficult subsets of the genome or when using variant calling methods that produce many complex variant calls, these corner cases can become relevant. Moreover, when benchmarking against gold-standard datasets that cover difficult regions of the genome (e.g. Platinum Genomes), the more complicated subsets of the genome will be responsible for most of the difference between methods.
Another component of hap.py is a variant pre-processing method which deals with complex variant representations and MNPs. When different callers may represent variants using a different number of VCF records, we should attempt to count these in a consistent fashion between methods. One example is the representation of MNVs as individual SNPs vs. as complex variants.
Consider the following case:
Complex variant representation:
chrQ 16 GGG TTT 0/1
vs.
Atomized representation:
chrQ 16 G T 0/1
chrQ 17 G T 0/1
chrQ 18 G T 0/1
If this variant is a false-positive, the first representation would naively contribute a single FP record. A variant caller that outputs the second representation would instead receive a penalty of three FPs for making the same variant call. Overall, the difference between the two representations might show significantly when looking at precision levels or false-positive rates (since these are relative to the total number of query counts, which use the same representations), but become important when we need to compare absolute numbers of false-positives. For this case, hap.py can perform a re-alignment of REF and ALT alleles on the query VCF, and splits the records into atomic variant alleles to produce more granular counts using pre.py. Left-shifting and trimming are also supported.
Hap.py includes a module to produce stratified variant counts. Variant types are determined using a re-alignment of REF and ALT alleles. This is more reliable than only using allele lengths. Consider the following complex deletion.
chr1 201586350 . CTCTCTCTCT CA
This complex variant call is equivalent to a deletion, followed by a SNP. Our quantification code will recognize this variant as a deletion and a SNP, and will count it in both categories (so a TP call for this variant will contribute a SNP and an INDEL). This effectively deals with variant calling methods that prefer to combine local haplotypes in the same variant records (e.g. Freebayes / Platypus), which would otherwise fall into a hard-to-assess "COMPLEX" variant call category that varies substantially between different variant calling methods.
chr1 201586350 . CTCTCTCTC C
chr1 201586359 . T A
Another feature of the quantification module in hap.py is stratification into variant sub-types and into genomic regions. For example, precision and recall can be computed at the same time for all GA4GH stratification regions, and for different INDEL lengths (<5, 7-15, 16+). Hap.py also calculates het-hom and Ti/Tv ratios for all subsets of benchmarked variants. Note that all region matching in hap.py is based on reference coordinates only. One case where this can lead to counterintuitive results is when considering hompolymer insertions:
Reference:
>chrQ
CAAAAA
VCF:
chrQ 1 C CA 0/1
BED for homopolymers:
1 6
In this example, the variant call given above would not be captured by the bed region for the homopolymers because it is associated with the reference base just before. To account for this, the bed intervals need to be expanded to include the padding base just before the regions.
Finally, we produce input data for ROC and precision/recall curves. An example is included.
The main two tools are hap.py (diploid precision/recall evaluation) and som.py (somatic precision/recall evaluation -- this ignores the GT and just checks for presence of alleles). Other tools are qfy.py (which just executes the quantification step of the analysis pipeline, this requires a GA4GH-intermediate VCF file), and pre.py, which is hap.py's input cleaning and variant normalisation step.
Here are some small example command lines. Advanced features like confident call / ambiguity / FP regions are also available, see the documentation for each tool for these.
Below, we assume that the code has been installed to the directory ${HAPPY}.
See also doc/happy.md.
$ ${HAPPY}/bin/hap.py \
example/happy/PG_NA12878_chr21.vcf.gz \
example/happy/NA12878_chr21.vcf.gz \
-f example/happy/PG_Conf_chr21.bed.gz \
-o test
$ ls test.*
test.metrics.json test.summary.csvThis example compares an example run of GATK 1.6 on NA12878 agains the Platinum Genomes reference dataset (Note: this is a fairly old version of GATK, so don't rely on these particular numbers for competitive comparisons!).
The summary CSV file contains all high-level metrics:
| Type | TRUTH.TOTAL | QUERY.TOTAL | METRIC.Recall | METRIC.Precision | METRIC.Frac_NA | TRUTH.TOTAL.TiTv_ratio | QUERY.TOTAL.TiTv_ratio | TRUTH.TOTAL.het_hom_ratio | QUERY.TOTAL.het_hom_ratio |
|---|---|---|---|---|---|---|---|---|---|
| INDEL | 9124 | 9905 | 0.869406 | 0.978441 | 0.194548 | NaN | NaN | 1.463852 | 1.209105 |
| SNP | 52520 | 48078 | 0.894478 | 0.998258 | 0.021070 | 2.081002 | 2.082603 | 1.595621 | 1.487599 |
These numbers tell us the SNP and indel recall of our query VCF against the truth dataset. See doc/happy.md for more documentation and some advice for their interpretation.
Som.py is a simple comparison tool based on bcftools. It does not perform genotype or haplotype matching.
See doc/sompy.md for more documentation.
${HAPPY}/bin/som.py example/sompy/PG_admix_truth_snvs.vcf.gz \
example/sompy/strelka_admix_snvs.vcf.gz \
-f example/sompy/FP_admix.bed.gz \
-o test
[...]
type total.truth total.query tp fp fn unk ambi recall recall2 precision na ambiguous
1 SNVs 16235 47530 15573 14698 662 17259 0 0.959224 0.959224 0.514453 0.363118 0
3 records 16235 47779 15573 14737 662 17469 0 0.959224 0.959224 0.513791 0.365621 0
ls test.*
test.stats.csv
The most relevant metrics are recall and precision. UNK calls are the calls that are outside the coverage of the truthset.
The simplest way to install hap.py is to use the helper script and your system Python install (requires these packages: Cython, Scipy, numpy, pandas, pybedtools, pysam, bx-python).
This command installs everything into ~/hap.py-install:
python install.py ~/hap.py-install
To also download rtgtools during the installation process and deploy it with this version
of hap.py, you can add the --with-rtgtools flag. For this to work, you must have a working
installation of Java 1.8 and Apache Ant 1.9.x.
python install.py ~/hap.py-install --with-rtgtools
The installer has an option --boost-root that allows us to use a specific installation of boost
(see above for instructions):
python install.py ~/hap.py-install --boost-root $HOME/boost_1_55_0_install
To use a special version of Python, run the installer with it:
$HOME/my-virtualenv/bin/python install.py ~/hap.py-install
To create a virtualenv:
python install.py ~/workspace-is/hap.py-install --python=virtualenv --python-virtualenv-dir=$HOME/my-virtualenv/hc.ve
There are various workaround / testing switches:
--python-virtualenv-updateupdates an existing virtualenv--python-virtualenv-forceoverwrites the virtualenv if it exists--pip-fix-certworks around outdated SSL certificates when using pip--no-testsdisables the unit/integration tests after installation--no-rebuild-externaldon't rebuild the external dependencies (htslib, ...) unless necessary--sge-moderequire switch--force-interactiveto run hap.py interactively (useful to prevent running on a head node when installing on systems with SGE)
Clone this repository and build a Docker image as follows.
$ sudo docker build .
$ sudo docker images
REPOSITORY TAG IMAGE ID CREATED VIRTUAL SIZE
<...> latest 3d03a99b3d81 1 second ago <...>
$ sudo docker run -ti --rm 3d03a99b3d81 bin/bash
$/ /opt/hap.py/bin/hap.py
A pre-built docker image can be found here: https://hub.docker.com/r/pkrusche/hap.py. It can be obtained by running:
docker pull pkrusche/hap.pyIf the current directory contains a clone of the hap.py repository, hap.py can be run in Docker as follows:
sudo docker run -it -v `pwd`:/data pkrusche/hap.py /opt/hap.py/bin/hap.py /data/example/PG_performance.vcf.gz /data/example/performance.vcf.gz -o /data/testThe -v argument mounts the current directory as /data in the Docker image. The output should also
appear in the current directory.
The default Docker image is based on Ubuntu. To use a Centos6 image as a base, use Dockerfile.centos6.
docker build -f Dockerfile.centos6 .
You will need these tools / libraries on your system to compile the code:
- CMake > 2.8
- GCC/G++ 4.9.2+ for compiling
- Boost 1.55+
- Python 2, version 2.7.8 or greater
- Python packages: Pandas, Numpy, Scipy, pysam, bx-python
- Java 1.8 when using vcfeval.
Then to compile:
- Get a hap.py checkout:
git clone https://github.com/sequencing/hap.py
- Make a build folder
mkdir hap.py-build cd hap.py-build - Run CMake
cmake ../hap.py
- Build
make
If this is successful, the bin subdirectory of your build folder will contain binaries and scripts:
$ python bin/hap.py --version
Hap.py v0.3.7Note that hap.py will copy all Python source files to the build folder, so when making changes to
any Python component, make must be run to make sure the scripts in the build folder are
up-to-date.
The source for hap.py contains a script configure.sh which shows some basic additional configuration flags, and an automated way to pre-package CMake setups. Here is a list of additional flags for CMake to change compile options help it find dependencies:
-DCMAKE_BUILD_TYPE=Debug-- set the build type, allowed values areDebugandRelease-DCMAKE_C_COMPILER=/usr/bin/gccand-DCMAKE_CXX_COMPILER=/usr/bin/g++-- change the compiler path-DCMAKE_INSTALL_PREFIX=/usr/local-- set an installation directory that will be used by make install.-DBOOST_ROOT=$HOME/boost_1_55_0_install-- set the path to Boost. Run the following commands to compile and install boost:
cd ~
wget http://downloads.sourceforge.net/project/boost/boost/1.55.0/boost_1_55_0.tar.bz2
tar xjf boost_1_55_0.tar.bz2
cd boost_1_55_0
./bootstrap.sh --with-libraries=filesystem,chrono,thread,iostreams,system,regex,test,program_options
./b2 --prefix=$HOME/boost_1_55_0_install install-DUSE_SGE-- enable the--force-interactiveswitch in hap.py.-DBUILD_VCFEVAL=ON-- Download and build rtgtools / vcfeval. This is a comparison engine that can be used as an alternative to the built-in xcmp in hap.py. To successfully build and run vcfeval, you will need:- A Java JRE, newer than 1.8.x
- ant > 1.9.2 (older versions of ant will not successfully build rtgtools)
See src/sh/illumina-setup.sh for an example. If running Java requires any special setup
(or to configure any other environment variables), you can specify a wrapper script using
-DVCFEVAL_WRAPPER={absolute_path_to_wrapper_script}. See src/sh/rtg-wrapper.sh for an example.
Compiling and testing can be done on a standard desktop system with 8GB of RAM. Whole-genome
comparisons (e.g. comparing a gVCF file against the Platinum Genomes truth dataset)
can use up to 64GB of RAM (20GB typical, depending on the input VCF) and about 4-12 minutes
using 40 processor cores. Whole exome comparison (using an exome bed mask and the -T switch)
can be carried out on a desktop system.
Hap.py is known to build and run on the following linux distributions (see also the Dockerfile for a list of required packages):
Ubuntu 12.04,14.04,16.04
CentOS 5,6,7
Hap.py must be compiled with g++ version 4.9.x or later, or with a recent version of Clang (testing is performed with g++).
Hap.py builds and passes basic tests on OS X 10.9+, but full WGS analyses are not tested for this platform.
Hap.py is not tested on Windows. The main dependency that fails compilation is htslib. Given a build of htslib and pysam, using hap.py on Windows should be possible.
Hap.py requires a human genome reference sequence which contains at least chromosomes 1-22,X,Y,M. The chromosomes should be named chr1-chr22, chrX, chrY, chrM. there is a script in src/sh/make_hg19.sh to create such a sequence, but you can also specify your own. In order for the integration tests to run successfully, it is necessary to point hap.py to the reference sequence using
export HGREF=<path-to-hg19.fa>Note that, while the test cases are based on hg19, other reference sequences are usable as well once the tool is installed.
Hap.py also requires a copy of the Boost libraries to work, with version >= 1.55. If compilation should fail using the included version of boost, you can compile a subset of boost like this:
cd ~
wget http://downloads.sourceforge.net/project/boost/boost/1.55.0/boost_1_55_0.tar.bz2
tar xjf boost_1_55_0.tar.bz2
cd boost_1_55_0
./bootstrap.sh --with-libraries=filesystem,chrono,thread,iostreams,system,regex,test,program_options
./b2 --prefix=$HOME/boost_1_55_0_install installYou can point Cmake to your version of boost as follows:
export BOOST_ROOT=$HOME/boost_1_55_0_installThe complete list of dependencies / packages to install beforehand can be found in the Dockerfile.