Release notes






This release drop support for Python 2.7.


  • Fixed a bug in allel.GenotypeDaskArray.to_allele_counts() where the shape of the output array was not being determined correctly. By Nick Harding, #266.



Use of the allel.stats namespace is deprecated in this release, all functions from stats modules are available from the root allel namespace, please access them from there.


Python 2.7 has had a stay of execution - this release supports Python 2.7 and 3.5-3.7. However, support for Python 2.7 will definitely be removed in version 1.3.

  • Added a new function allel.pbs() which implements the Population Branching Statistic (PBS), a test for selection based on allel frequency differentiation between three populations. #210.
  • Added a new function allel.roh_poissonhmm() for inferring runs of homozygosity, which uses a Poisson HMM and is orders of magnitude faster than the previously vailable allel.roh_mhmm() multinomial HMM implementation. By Nick Harding, #188, #187.
  • Added a workaround to ensure arrays passed into Cython functions are safe to use as memoryviews, which is required to avoid errors when using distributed computing systems like dask.distributed where data may be moved between compute nodes and passed with a read-only flag set. #208, #206.
  • Added support for parsing VCF files where the chromosomes are not in lexical sorted order. Also improved handling of cases where no variants are returned. #221, #167, #213.
  • Added a new index class allel.ChromPosIndex for locating data given chromosome and positio locations. This behaves similarly to the existing allel.SortedMultiIndex but the chromosome values do not need to be sorted. #201, #239.
  • Added new parameters exclude_fields and rename_fields to VCF parsing functions to add greater flexibility when selecting fields to extract. Also added several measures to protect against name clashes when converting VCF to Zarr on platforms with a case-insensitive file system. #215, #216.
  • Added a convenience function allel.read_vcf_headers(), to obtain just header information from a VCF file. #216.
  • All functions for computing site frequency spectra now accept an optional argument n for manually specifying the number of chromosomes sampled from the population. #174, #240.
  • Added start, stop and step options to allel.equally_accessible_windows(). #234, #166.
  • Fixed broken implementation of allel.AlleleCountsArray.map_alleles(). #241, #200.
  • Fixed functions calculating Tajima’s D such that a value of np.nan is returned if there are fewer than 3 segregating sites. By Andrew Kern and Alistair Miles, #175, #237.
  • Fixed incorrect fill value in GFF parsing functions. #165, #223.
  • Fixed a problem in count_alleles() methods where a subpop arg was provided as a numpy array. #235, #171.
  • Removed fill option to LD functions allel.rogers_huff_r() and allel.rogers_huff_r_between(), always use NaN where a value cannot be calculated. Also added additional tests and check for case where variants have no data. #197, #243.
  • Allow multiallelic variants in allel.ehh_decay(). #209, #244.
  • Added checks to raise appropriate errors if user tries to rename two fields to the same name when reading VCF. #245, #220.
  • Fixed so that installation of numpy prior to installation of scikit-allel is no longer required - numpy will be automatically installed as a dependency if not already installed. By @haseley, #212, #211.
  • Migrate to using pytest instead of nose for testing. #236, #184.
  • Small optimisation for writing zarr attributes. #225, #238.
  • Fixed pandas deprecation warning. By Summer Rae, #228.
  • Fixed problem where some packages where getting clobbered by imports of other packages. #163, #232.
  • Added support for Python 3.7 and compatibility with numpy 1.15. #217, #214.
  • Various documentation improvements. By Peter Ralph and CJ Battey, #229.


  • Various VCF parsing improvements and bug fixes (#183, #189).


  • Added support for Type=Character in VCF files (Kunal Bhutani; #159)
  • Fixed type of indexing variables in VCF reading functions to handle larger datasets (#160).
  • Added option to specify string codec in allel.vcf_to_zarr() (#156).
  • Fixed bug in LD plotting function (#161).


  • Changed semantics of is_snp computed field when extracting data from VCF to exclude variants where one of the alternate alleles is a spanning deletion (‘*’) (#155).
  • Resolved minor logging bug (#152).


  • Added an option to allel.vcf_to_hdf5() to disable use of variable length strings because they can cause large HDF5 file size (#153).


  • Include fixture data in release to aid testing and binary builds.


Reading Variant Call Format (VCF) files

This release includes new functions for extracting data from VCF files and loading into NumPy arrays, HDF5 files and other storage containers. These functions are backed by VCF parsing code implemented in Cython, so should be reasonably fast. This is new code so there may be bugs, please report any issues via GitHub.

For a tutorial and worked examples, see the following article: Extracting data from VCF.

For API documentation, see the following functions: allel.read_vcf(), allel.vcf_to_npz(), allel.vcf_to_hdf5(), allel.vcf_to_zarr(), allel.vcf_to_dataframe(), allel.vcf_to_csv(), allel.vcf_to_recarray(), allel.iter_vcf_chunks().

Reading GFF3 files

Added convenience functions allel.gff3_to_dataframe() and allel.gff3_to_recarray().

Maintenance work

End of support for Python 2


This is the last version of scikit-allel that will support Python 2. The next version of scikit-allel will support Python versions 3.5 and later only.


Fix test compatibility with numpy 1.10.


Move cython function imports outside of functions to work around bug found when using scikit-allel with dask.


Add missing test packages so full test suite can be run to verify install.


This release includes some subtle but important changes to the architecture of the data structures modules (allel.model.ndarray, allel.model.chunked, allel.model.dask). These changes are mostly backwards-compatible but in some cases could break existing code, hence the major version number has been incremented. Also included in this release are some new functions related to Mendelian inheritance and calling runs of homozygosity, further details below.

Mendelian errors and phasing by transmission

This release includes a new allel.stats.mendel module with functions to help with analysis of related individuals. The function allel.mendel_errors() locates genotype calls within a trio or cross that are not consistent with Mendelian segregation of alleles. The function allel.phase_by_transmission() will resolve unphased diploid genotypes into phased haplotypes for a trio or cross using Mendelian transmission rules. The function allel.paint_transmission() can help with evaluating and visualizing the results of phasing a trio or cross.

Runs of homozygosity

A new allel.roh_mhmm() function provides support for locating long runs of homozygosity within a single sample. The function uses a multinomial hidden Markov model to predict runs of homozygosity based on the rate of heterozygosity over the genome. The function can also incorporate information about which positions in the genome are not accessible to variant calling and hence where there is no information about heterozygosity, to reduce false calling of ROH in regions where there is patchy data. We’ve run this on data from the Ag1000G project but have not performed a comprehensive evaluation with other species, feedback is very welcome.

Changes to data structures

The allel.model.ndarray module includes a new allel.model.ndarray.GenotypeVector class. This class represents an array of genotype calls for a single variant in multiple samples, or for a single sample at multiple variants. This class makes it easier, for example, to locate all variants which are heterozygous in a single sample.

Also in the same module are two new classes allel.model.ndarray.GenotypeAlleleCountsArray and allel.model.ndarray.GenotypeAlleleCountsVector. These classes provide support for an alternative encoding of genotype calls, where each call is stored as the counts of each allele observed. This allows encoding of genotype calls where samples may have different ploidy for a given chromosome (e.g., Leishmania) and/or where samples carry structural variation within some genome regions, altering copy number (and hence effective ploidy) with respect to the reference sequence.

There have also been architectural changes to all data structures modules. The most important change is that all classes in the allel.model.ndarray module now wrap numpy arrays and are no longer direct sub-classes of the numpy numpy.ndarray class. These classes still behave like numpy arrays in most respects, and so in most cases this change should not impact existing code. If you need a plain numpy array for any reason you can always use numpy.asarray() or access the .values property, e.g.:

>>> import allel
>>> import numpy as np
>>> g = allel.GenotypeArray([[[0, 1], [0, 0]], [[0, 2], [1, 1]]])
>>> isinstance(g, np.ndarray)
>>> a = np.asarray(g)
>>> isinstance(a, np.ndarray)
>>> isinstance(g.values, np.ndarray)

This change was made because there are a number of complexities that arise when sub-classing class:numpy.ndarray and these were proving tricky to manage and maintain.

The allel.model.chunked and allel.model.dask modules also follow the same wrapper pattern. For the allel.model.dask module this means a change in the way that classes are instantiated. For example, to create a allel.model.dask.GenotypeDaskArray, pass the underlying data directly into the class constructor, e.g.:

>>> import allel
>>> import h5py
>>> h5f = h5py.File('callset.h5', mode='r')
>>> h5d = h5f['3R/calldata/genotype']
>>> genotypes = allel.GenotypeDaskArray(h5d)

If the underlying data is chunked then there is no need to specify the chunks manually when instantiating a dask array, the native chunk shape will be used.

Finally, the allel.model.bcolz module has been removed, use either the allel.model.chunked or allel.model.dask module instead.


This release resolves compatibility issues with Zarr version 2.1.


  • Added parameter min_maf to allel.ihs() to skip IHS calculation for variants below a given minor allele frequency.
  • Minor change to calculation of integrated haplotype homozygosity to enable values to be reported for first and last variants if include_edges is True.
  • Minor change to allel.standardize_by_allele_count() to better handle missing values.


In this release the implementations of allel.ihs() and allel.xpehh() selection statistics have been reworked to address a number of issues:

  • Both functions can now integrate over either a genetic map (via the map_pos parameter) or a physical map.
  • Both functions now accept max_gap and gap_scale parameters to perform adjustments to integrated haplotype homozygosity where there are large gaps between variants, following the standard approach. Alternatively, if a map of genome accessibility is available, it may be provided via the is_accessible parameter, in which case the distance between variants will be scaled by the fraction of accessible bases between them.
  • Both functions are now faster and can make use of multiple threads to further accelerate computation.
  • Several bugs in the previous implementations of these functions have been fixed (#91).
  • New utility functions are provided for standardising selection scores, see allel.standardize_by_allele_count() (for use with IHS and NSL) and allel.standardize() (for use with XPEHH).

Other changes:


  • Fixed a bug in the count_alleles() methods on genotype and haplotype array classes that manifested if the max_allele argument was provided (#59).
  • Fixed a bug in Jupyter notebook display method for chunked tables (#57).
  • Fixed a bug in site frequency spectrum scaling functions (#54).
  • Changed behaviour of subset method on genotype and haplotype arrays to better infer argument types and handle None argument values (#55).
  • Changed table eval and query methods to make python the default for expression evaluation, because it is more expressive than numexpr (#58).




  • Added new allel.model.dask module, providing implementations of the genotype, haplotype and allele counts classes backed by dask.array (#32).
  • Released the GIL where possible in Cython optimised functions (#43).
  • Changed functions in allel.stats.selection that accept min_ehh argument, such that min_ehh = None should now be used to indicate that no minimum EHH threshold should be applied.


The major change in v0.19.0 is the addition of the new allel.model.chunked module, which provides classes for variant call data backed by chunked array storage (#31). This is a generalisation of the previously available allel.model.bcolz to enable the use of both bcolz and HDF5 (via h5py) as backing storage. The allel.model.bcolz module is now deprecated but will be retained for backwargs compatibility until the next major release.

Other changes:

Contributors: alimanfoo, hardingnj


  • Minor change to the Garud H statistics to avoid raising an exception when the number of distinct haplotypes is very low (#20).



  • Added new module for computing and plotting site frequency spectra, see allel.stats.sf (#12).
  • All plotting functions have been moved into the appropriate stats module that they naturally correspond to. The allel.plot module is deprecated (#13).
  • Improved performance of carray and ctable loading from HDF5 with a condition (#11).


  • Fixed behaviour of take() method on compressed arrays when indices are not in increasing order (#6).
  • Minor change to scaler argument to PCA functions in allel.stats.decomposition to avoid confusion about when to fall back to default scaler (#7).



  • Added new selection module with functions for haplotype-based analyses of recent selection, see allel.stats.selection.


  • Improved performance of allel.model.bcolz.carray_block_compress(), allel.model.bcolz.ctable_block_compress() and allel.model.bcolz.carray_block_subset() for very sparse selections.
  • Fix bug in IPython HTML table captions.
  • Fix bug in addcol() method on bcolz ctable wrappers.


  • Fix missing package in


  • Added functions to estimate Fst with standard error via a block-jackknife: allel.blockwise_weir_cockerham_fst(), allel.blockwise_hudson_fst(), allel.blockwise_patterson_fst().
  • Fixed a serious bug in allel.weir_cockerham_fst() related to incorrect estimation of heterozygosity, which manifested if the subpopulations being compared were not a partition of the total population (i.e., there were one or more samples in the genotype array that were not included in the subpopulations to compare).
  • Added method allel.AlleleCountsArray.max_allele() to determine highest allele index for each variant.
  • Changed first return value from admixture functions allel.blockwise_patterson_f3() and allel.blockwise_patterson_d() to return the estimator from the whole dataset.
  • Added utility functions to the allel.stats.distance module for transforming coordinates between condensed and uncondensed forms of a distance matrix.
  • Classes previously available from the allel.model and allel.bcolz modules are now aliased from the root allel module for convenience. These modules have been reorganised into an allel.model package with sub-modules allel.model.ndarray and allel.model.bcolz.
  • All functions in the allel.model.bcolz module use cparams from input carray as default for output carray (convenient if you, e.g., want to use zlib level 1 throughout).
  • All classes in the allel.model.ndarray and allel.model.bcolz modules have changed the default value for the copy keyword argument to False. This means that not copying the input data, just wrapping it, is now the default behaviour.
  • Fixed bug in GenotypeArray.to_gt() where maximum allele index is zero.


  • Added a new module allel.stats.admixture with statistical tests for admixture between populations, implementing the f2, f3 and D statistics from Patterson (2012). Functions include allel.blockwise_patterson_f3() and allel.blockwise_patterson_d() which compute the f3 and D statistics respectively in blocks of a given number of variants and perform a block-jackknife to estimate the standard error.


  • Added functions for principal components analysis of genotype data. Functions in the new module allel.stats.decomposition include allel.pca() to perform a PCA via full singular value decomposition, and allel.randomized_pca() which uses an approximate truncated singular value decomposition to speed up computation. In tests with real data the randomized PCA is around 5 times faster and uses half as much memory as the conventional PCA, producing highly similar results.
  • Added function allel.pcoa() for principal coordinate analysis (a.k.a. classical multi-dimensional scaling) of a distance matrix.
  • Added new utility module allel.stats.preprocessing with classes for scaling genotype data prior to use as input for PCA or PCoA. By default the scaling (i.e., normalization) of Patterson (2006) is used with principal components analysis functions in the allel.stats.decomposition module. Scaling functions can improve the ability to resolve population structure via PCA or PCoA.
  • Added method allel.GenotypeArray.to_n_ref(). Also added dtype argument to allel.GenotypeArray.to_n_ref() and allel.GenotypeArray.to_n_alt() methods to enable direct output as float arrays, which can be convenient if these arrays are then going to be scaled for use in PCA or PCoA.
  • Added allel.GenotypeArray.mask property which can be set with a Boolean mask to filter genotype calls from genotype and allele counting operations. A similar property is available on the allel.GenotypeCArray class. Also added method allel.GenotypeArray.fill_masked() and similar method on the allel.GenotypeCArray class to fill masked genotype calls with a value (e.g., -1).




  • Added documentation for the functions allel.bcolz.carray_from_hdf5(), allel.bcolz.carray_to_hdf5(), allel.bcolz.ctable_from_hdf5_group(), allel.bcolz.ctable_to_hdf5_group().
  • Refactoring of internals within the allel.bcolz module.



  • Added function allel.write_fasta() for writing a nucleotide sequence stored as a NumPy array out to a FASTA format file.