def _finish(self, default):
assert len(self._cases) > 0
from hail.expr.functions import cond
expr = default
for conditional, then in self._cases[::-1]:
expr = cond(conditional, then, expr, missing_false=self._missing_false)
return expr
def _finish(self, default):
assert len(self._cases) > 0
from hail.expr.functions import cond
expr = default
for conditional, then in self._cases[::-1]:
expr = cond(conditional, then, expr, missing_false=self._missing_false)
return expr
def hwe_normalized_pca(dataset, k=10, compute_loadings=False, as_array=False):
"""Run principal component analysis (PCA) on the Hardy-Weinberg-normalized call matrix.
Examples
--------
>>> eigenvalues, scores, loadings = methods.hwe_normalized_pca(dataset, k=5)
Notes
-----
Variants that are all homozygous reference or all homozygous variant are removed before evaluation.
Parameters
----------
dataset : :class:`.MatrixTable`
Dataset.
k : :obj:`int`
Number of principal components.
compute_loadings : :obj:`bool`
If ``True``, compute row loadings.
as_array : :obj:`bool`
If ``True``, return scores and loadings as an array field. If ``False``, return
one field per element (`PC1`, `PC2`, ... `PCk`).
Returns
-------
(:obj:`list` of :obj:`float`, :class:`.Table`, :class:`.Table`)
List of eigenvalues, table with column scores, table with row loadings.
"""
dataset = dataset.annotate_rows(AC=agg.sum(dataset.GT.num_alt_alleles()),
n_called=agg.count_where(
functions.is_defined(dataset.GT)))
dataset = dataset.filter_rows(
(dataset.AC > 0) & (dataset.AC < 2 * dataset.n_called)).persist()
n_variants = dataset.count_rows()
if n_variants == 0:
raise FatalError(
"Cannot run PCA: found 0 variants after filtering out monomorphic sites."
)
info("Running PCA using {} variants.".format(n_variants))
entry_expr = functions.bind(
dataset.AC / dataset.n_called, lambda mean_gt: functions.cond(
functions.is_defined(dataset.GT), (dataset.GT.num_alt_alleles(
) - mean_gt) / functions.sqrt(mean_gt *
(2 - mean_gt) * n_variants / 2), 0))
result = pca(entry_expr, k, compute_loadings, as_array)
dataset.unpersist()
return result
def f(base):
# build cond chain bottom-up
expr = default
for condition, then in self._cases[::-1]:
expr = cond(condition, then, expr)
return expr
def grm(dataset):
"""Compute the Genetic Relatedness Matrix (GRM).
.. include:: ../_templates/req_tvariant.rst
.. include:: ../_templates/req_biallelic.rst
Examples
--------
>>> km = methods.grm(dataset)
Notes
-----
The genetic relationship matrix (GRM) :math:`G` encodes genetic correlation
between each pair of samples. It is defined by :math:`G = MM^T` where
:math:`M` is a standardized version of the genotype matrix, computed as
follows. Let :math:`C` be the :math:`n \\times m` matrix of raw genotypes
in the variant dataset, with rows indexed by :math:`n` samples and columns
indexed by :math:`m` bialellic autosomal variants; :math:`C_{ij}` is the
number of alternate alleles of variant :math:`j` carried by sample
:math:`i`, which can be 0, 1, 2, or missing. For each variant :math:`j`,
the sample alternate allele frequency :math:`p_j` is computed as half the
mean of the non-missing entries of column :math:`j`. Entries of :math:`M`
are then mean-centered and variance-normalized as
.. math::
M_{ij} = \\frac{C_{ij}-2p_j}{\sqrt{2p_j(1-p_j)m}},
with :math:`M_{ij} = 0` for :math:`C_{ij}` missing (i.e. mean genotype
imputation). This scaling normalizes genotype variances to a common value
:math:`1/m` for variants in Hardy-Weinberg equilibrium and is further
motivated in the paper `Patterson, Price and Reich, 2006
<http://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.0020190>`__.
(The resulting amplification of signal from the low end of the allele
frequency spectrum will also introduce noise for rare variants; common
practice is to filter out variants with minor allele frequency below some
cutoff.) The factor :math:`1/m` gives each sample row approximately unit
total variance (assuming linkage equilibrium) so that the diagonal entries
of the GRM are approximately 1. Equivalently,
.. math::
G_{ik} = \\frac{1}{m} \\sum_{j=1}^m \\frac{(C_{ij}-2p_j)(C_{kj}-2p_j)}{2 p_j (1-p_j)}
Warning
-------
Since Hardy-Weinberg normalization cannot be applied to variants that
contain only reference alleles or only alternate alleles, all such variants
are removed prior to calcularing the GRM.
Parameters
----------
dataset : :class:`.MatrixTable`
Dataset to sample from.
Returns
-------
:class:`genetics.KinshipMatrix`
Genetic Relatedness Matrix for all samples.
:rtype:
"""
dataset = dataset.annotate_rows(AC=agg.sum(dataset.GT.num_alt_alleles()),
n_called=agg.count_where(
functions.is_defined(dataset.GT)))
dataset = dataset.filter_rows(
(dataset.AC > 0) & (dataset.AC < 2 * dataset.n_called)).persist()
n_variants = dataset.count_rows()
if n_variants == 0:
raise FatalError(
"Cannot run GRM: found 0 variants after filtering out monomorphic sites."
)
info("Computing GRM using {} variants.".format(n_variants))
normalized_genotype_expr = functions.bind(
dataset.AC / dataset.n_called, lambda mean_gt: functions.cond(
functions.is_defined(dataset.GT), (dataset.GT.num_alt_alleles(
) - mean_gt) / functions.sqrt(mean_gt *
(2 - mean_gt) * n_variants / 2), 0))
bm = BlockMatrix.from_matrix_table(normalized_genotype_expr)
dataset.unpersist()
grm = bm.T.dot(bm)
return KinshipMatrix._from_block_matrix(
dataset.colkey_schema, grm,
[row.s
for row in dataset.cols_table().select('s').collect()], n_variants)
def f(base):
# build cond chain bottom-up
expr = default
for condition, then in self._cases[::-1]:
expr = cond(condition, then, expr)
return expr
