Reading VCF#

The function converts one or more VCF files to Zarr files stored in sgkit’s Xarray data representation.


  • Reads bgzip-compressed VCF and BCF files.

  • Large VCF files can be partitioned into regions using a Tabix (.tbi) or CSI (.csi) index, and each region is processed in parallel using Dask.

  • VCF parsing is performed by cyvcf2, a Cython wrapper around htslib, the industry-standard VCF library.

  • Control over Zarr chunk sizes allows VCFs with a large number of samples to be converted efficiently.

  • Input and output files can reside on local filesystems, Amazon S3, or Google Cloud Storage.

  • Support for polyploid and mixed-ploidy genotypes.


VCF support is an “extra” feature within sgkit and requires additional dependencies, notably cyvcf2.

To install sgkit with VCF support using pip (there is no conda package):

$ pip install 'sgkit[vcf]'

There are installation instructions for cyvcf2, which may be helpful if you encounter errors during installation.


Reading VCFs is not supported on Windows, since cyvcf2 and htslib do not currently work on Windows. As a workaround, consider using scikit-allel’s vcf_to_zarr function to write a VCF in Zarr format, followed by sgkit.read_scikit_allel_vcfzarr() to read the VCF as a xarray.Dataset.


To convert a single VCF or BCF file to Zarr, just specify the input and output file names:

>>> import sgkit as sg
>>> from import vcf_to_zarr
>>> vcf_to_zarr("CEUTrio.20.21.gatk3.4.g.vcf.bgz", "output.zarr")
>>> ds = sg.load_dataset("output.zarr")
>>> ds
Dimensions:               (alleles: 4, ploidy: 2, samples: 1, variants: 19910)
Dimensions without coordinates: alleles, ploidy, samples, variants
Data variables:
    call_genotype         (variants, samples, ploidy) int8 dask.array<chunksize=(10000, 1, 2), meta=np.ndarray>
    call_genotype_mask    (variants, samples, ploidy) bool dask.array<chunksize=(10000, 1, 2), meta=np.ndarray>
    call_genotype_phased  (variants, samples) bool dask.array<chunksize=(10000, 1), meta=np.ndarray>
    sample_id             (samples) <U7 dask.array<chunksize=(1,), meta=np.ndarray>
    variant_allele        (variants, alleles) object dask.array<chunksize=(10000, 4), meta=np.ndarray>
    variant_contig        (variants) int8 dask.array<chunksize=(10000,), meta=np.ndarray>
    variant_id            (variants) object dask.array<chunksize=(10000,), meta=np.ndarray>
    variant_id_mask       (variants) bool dask.array<chunksize=(10000,), meta=np.ndarray>
    variant_position      (variants) int32 dask.array<chunksize=(10000,), meta=np.ndarray>
    contigs:                    ['20', '21']
    max_variant_allele_length:  48
    max_variant_id_length:      1

The function can accept multiple files, and furthermore, each of these files can be partitioned to enable parallel processing.

Multiple files#

If there are multiple files, then pass a list:

>>> from import vcf_to_zarr
>>> vcf_to_zarr(["CEUTrio.20.gatk3.4.g.vcf.bgz", "CEUTrio.21.gatk3.4.g.vcf.bgz"], "output.zarr")

Processing multiple inputs is more work than a single file, since behind the scenes each input is converted to a separate temporary Zarr file on disk, then these files are concatenated and rechunked to form the final output Zarr file.

In the single file case, the input VCF is converted to the output Zarr file in a single sequential pass with no need for intermediate temporary files. For small files this is fine, but for very large files it’s a good idea to partition them so the conversion runs faster.


Partitioning a large VCF file involves breaking it into a number of roughly equal-sized parts that can be processed in parallel. The parts are specified using genomic regions that follow the regions format used in bcftools: chr:beg-end, where positions are 1-based and inclusive.

All files to be partitioned must have either a Tabix (.tbi) or CSI (.csi) index. If both are present for a particular file, then Tabix is used for finding partitions.

The function will create a list of region strings for a VCF file, given a desired number of parts to split the file into:

>>> from import partition_into_regions
>>> partition_into_regions("CEUTrio.20.21.gatk3.4.g.vcf.bgz", num_parts=10)
['20:1-10108928', '20:10108929-10207232', '20:10207233-', '21:1-10027008', '21:10027009-10043392', '21:10043393-10108928', '21:10108929-10141696', '21:10141697-10174464', '21:10174465-10190848', '21:10190849-10207232', '21:10207233-']

It’s important to note that the number of regions returned may not be exactly the number of parts requested: it may be more or less. However, it is guaranteed that the regions will be contiguous and will cover the whole VCF file.

The region strings are passed to vcf_to_zarr so it can process the parts in parallel:

>>> from import partition_into_regions, vcf_to_zarr
>>> regions = partition_into_regions("CEUTrio.20.21.gatk3.4.g.vcf.bgz", num_parts=10)
>>> vcf_to_zarr("CEUTrio.20.21.gatk3.4.g.vcf.bgz", "output.zarr", regions=regions)

It’s also possible to produce parts that have an approximate target size (in bytes). This is useful if you are partitioning multiple files, and want all the parts to be roughly the same size.

>>> from import partition_into_regions, vcf_to_zarr
>>> inputs = ["CEUTrio.20.gatk3.4.g.vcf.bgz", "CEUTrio.21.gatk3.4.g.vcf.bgz"]
>>> regions = [partition_into_regions(input, target_part_size=100_000) for input in inputs]
>>> vcf_to_zarr(inputs, "output.zarr", regions=regions)

The same result can be obtained more simply by specifying target_part_size in the call to vcf_to_zarr:

>>> from import vcf_to_zarr
>>> inputs = ["CEUTrio.20.gatk3.4.g.vcf.bgz", "CEUTrio.21.gatk3.4.g.vcf.bgz"]
>>> vcf_to_zarr(inputs, "output.zarr", target_part_size=100_000)

As a special case, None is used to represent a single partition.

>>> from import partition_into_regions
>>> partition_into_regions("CEUTrio.20.21.gatk3.4.g.vcf.bgz", num_parts=1)

Chunk sizes#

One key advantage of using Zarr as a storage format is its ability to store large files in chunks, making it straightforward to process the data in parallel.

You can control the chunk length (in the variants dimension) and width (in the samples dimension) by setting the chunk_length and chunk_width parameters to

Due to the way that VCF files are parsed, all of the sample data for a given chunk of variants are loaded into memory at one time. In other words, chunk_length is honored at read time, whereas chunk_width is honored at write time. For files with very large numbers of samples, this can exceed working memory. The solution is to also set temp_chunk_length to be a smaller number (than chunk_length), so that fewer variants are loaded into memory at one time, while still having the desired output chunk sizes (chunk_length and chunk_width). Note that temp_chunk_length must divide chunk_length evenly.

Cloud storage#

VCF files can be read from various file systems including cloud stores. However, since different underlying libraries are used in different functions, there are slight differences in configuration that are outlined here.

The function uses fsspec to read VCF metadata and their indexes. Therefore, to access files stored on Amazon S3 or Google Cloud Storage install the s3fs or gcsfs Python packages, and use s3:// or gs:// URLs.

You can also pass storage_options to to configure the fsspec backend. This provides a way to pass any credentials or other necessary arguments needed to s3fs or gcsfs.

The function does not use fsspec, since it relies on htslib for file handling, and therefore has its own way of accessing cloud storage. You can access files stored on Amazon S3 or Google Cloud Storage using s3:// or gs:// URLs. Setting credentials or other options is typically achieved using environment variables for the underlying cloud store.


Zarr offers a lot of flexibility over controlling how data is compressed. Each variable can use a different compression algorithm, and its own list of filters.

The function tries to choose good defaults for compression, using information about the variable’s dtype, and also the nature of the data being stored.

For example, variant_position (from the VCF POS field) is a monotonically increasing integer (within a contig) so it benefits from using a delta encoding to store the differences in its values, since these are smaller integers that compress better. This encoding is specified using the NumCodecs Delta codec as a Zarr filter.

When converting from VCF you can specify the default compression algorithm to use for all variables by specifying compressor in the call to There are trade-offs between compression speed and size, which this benchmark does a good job of exploring.

Sometimes you may want to override the compression for a particular variable. A good example of this is for VCF FORMAT fields that are floats. Floats don’t compress well, and since there is a value for every sample they can take up a lot of space. In many cases full float precision is not needed, so it is a good idea to use a filter to transform the float to an int, that takes less space.

For example, the following code creates an encoding that can be passed to to store the VCF DS FORMAT field to 2 decimal places. (DS is a dosage field that is between 0 and 2 so we know it will fit into an unsigned 8-bit int.):

from numcodecs import FixedScaleOffset

encoding = {
    "call_DS": {
        "filters": [FixedScaleOffset(offset=0, scale=100, dtype="f4", astype="u1")],

Note that this encoding won’t work for floats that may be NaN. Consider using Quantize (with astype=np.float16) or Bitround in that case.

Low-level operation#

Calling runs a two-step operation:

  1. Write the output for each input region to a separate temporary Zarr store

  2. Concatenate and rechunk the temporary stores into the final output Zarr store

Each step is run as a Dask computation, which means you can use any Dask configuration mechanisms to control aspects of the computation.

For example, you can set the Dask scheduler to run on a cluster. In this case you would set the temporary Zarr store to be a cloud storage URL (by setting tempdir) so that all workers can access the store (both for reading and writing).

For debugging, or for more control over the steps, consider using followed by

Polyploid and mixed-ploidy VCF#

The function can be used to convert polyploid VCF data to Zarr files stored in sgkit’s Xarray data representation by specifying the ploidy of the dataset using the ploidy parameter.

By default, sgkit expects VCF files to have a consistent ploidy level across all samples and variants. Manual specification of ploidy is necessary because, within the VCF standard, ploidy is indicated by the length of each genotype call and hence it may not be consistent throughout the entire VCF file.

If a genotype call of lower than specified ploidy is encountered it will be treated as an incomplete genotype. For example, if a VCF is being processed assuming a ploidy of four (i.e. tetraploid) then the diploid genotype 0/1 will be treated as the incomplete tetraploid genotype 0/1/./..

If a genotype call of higher than specified ploidy is encountered an exception is raised. This exception can be avoided using the truncate_calls parameter in which case the additional alleles will be skipped.

Conversion of mixed-ploidy VCF files is also supported by by use of the mixed_ploidy parameter. In this case ploidy specifies the maximum allowed ploidy and lower ploidy genotype calls within the VCF file will be preserved within the resulting dataset.

Note that many statistical genetics methods available for diploid data are not generalized to polyploid and or mixed-ploidy data. Therefore, some methods available in sgkit may only be applicable to diploid or fixed-ploidy datasets.

Methods that are generalized to polyploid and mixed-ploidy data may make assumptions such as polysomic inheritance and hence it is necessary to understand the type of polyploidy present within any given dataset.

Example: converting 1000 genomes VCF to Zarr#

This section shows how to convert the 1000 genomes dataset into Zarr format for analysis in sgkit.

For reference, the conversion (not including downloading the data) took about an hour on a machine with 32 vCPUs and 128GB of memory (GCP e2-standard-32).

Install sgkit#

Install the main package using conda or pip, and the VCF extra package using pip, as described in Installation.

Download the data#

Run the following to download the 1000 genomes VCF files over FTP:

mkdir -p data/1kg
for contig in {1..22}; do
  wget -P data/1kg${contig}.phase3_shapeit2_mvncall_integrated_v5b.20130502.genotypes.vcf.gz
  wget -P data/1kg${contig}.phase3_shapeit2_mvncall_integrated_v5b.20130502.genotypes.vcf.gz.tbi

Run the conversion#

Run the following Python code:

from import vcf_to_zarr
from dask.distributed import Client

if __name__ == "__main__":
    client = Client(n_workers=16, threads_per_worker=1)

    vcfs = [f"data/1kg/ALL.chr{contig}.phase3_shapeit2_mvncall_integrated_v5b.20130502.genotypes.vcf.gz" for contig in range(1, 23)]
    target = "1kg.zarr"
    vcf_to_zarr(vcfs, target, tempdir="1kg-tmp")

A few notes about the code:

1. Using a Dask distributed cluster, even on a single machine, performs better than the default scheduler (which uses threads), or the multiprocessing scheduler. Creating a Client object will start a local cluster.

2. Making the number of workers less than the number of cores (16 rather than 32 in this case) will improve performance. It’s also important to set threads_per_worker to 1 to avoid overcommitting threads, as recommended in the Dask documentation.

3. It is useful to track the progress of the computation using the Dask dashboard. There are two steps in the conversion operation, described in Low-level operation, the first of which has coarse-grained, long-running tasks, and the second which has much shorter-running tasks. There is a considerable delay (around 10 minutes) between the two steps, so don’t worry if it doesn’t look like it’s progressing.

4. Only the core VCF fields and genotypes are converted. To import more VCF fields see the documentation for the fields and field_defs parameters for

Inspect the dataset#

When the conversion is complete, have a look at the dataset as follows:

>>> import sgkit as sg
>>> ds = sg.load_dataset("1kg.zarr")
>>> ds
Dimensions:               (variants: 81271745, samples: 2504, ploidy: 2, alleles: 4)
Dimensions without coordinates: variants, samples, ploidy, alleles
Data variables:
    call_genotype         (variants, samples, ploidy) int8 dask.array<chunksize=(10000, 1000, 2), meta=np.ndarray>
    call_genotype_mask    (variants, samples, ploidy) bool dask.array<chunksize=(10000, 1000, 2), meta=np.ndarray>
    call_genotype_phased  (variants, samples) bool dask.array<chunksize=(10000, 1000), meta=np.ndarray>
    sample_id             (samples) object dask.array<chunksize=(1000,), meta=np.ndarray>
    variant_allele        (variants, alleles) object dask.array<chunksize=(10000, 4), meta=np.ndarray>
    variant_contig        (variants) int8 dask.array<chunksize=(10000,), meta=np.ndarray>
    variant_id            (variants) object dask.array<chunksize=(10000,), meta=np.ndarray>
    variant_id_mask       (variants) bool dask.array<chunksize=(10000,), meta=np.ndarray>
    variant_position      (variants) int32 dask.array<chunksize=(10000,), meta=np.ndarray>
    contigs:               ['1', '2', '3', '4', '5', '6', '7', '8', '9', '10'...
    max_alt_alleles_seen:  12