def build_pairwise_matrix(self, strain_names, elem_intervals):
# 3d matrix. First index is combo, remaining 2d matrices are counts for pairwise intervals
source_counts = np.zeros([(subspecies.NUM_SUBSPECIES + 1) ** 2, len(elem_intervals), len(elem_intervals)],
dtype=np.int16)
for strain_name in strain_names:
intervals, sources = self.sample_dict[strain_name]
# map this strain's intervals onto the elementary intervals
breaks = np.insert(np.searchsorted(elem_intervals, intervals), 0, -1)
for row in xrange(len(intervals)):
for col in xrange(row, len(intervals)): # only upper triangle
source = subspecies.combine(sources[row], sources[col])
source_ordinate = subspecies.to_ordinal(source)
source_counts[source_ordinate, breaks[row] + 1:breaks[row + 1] + 1,
breaks[col] + 1:breaks[col + 1] + 1] += 1
return source_counts
def interlocus_dependence(self, strain_names):
""" Performs a chi square test to find interval pairs whose origins are interdependent
:param strain_names: list of strain names to analyze
:return: elementary intervals, matrix of chi square values, matrix of p values (both upper triangular)
"""
combo_count_dict, intervals = self.pairwise_frequencies(strain_names)
# convert source_counts to matrix combo_counts
combo_counts = np.empty([len(intervals), len(intervals), subspecies.NUM_SUBSPECIES ** 2], dtype=np.uint16)
species_counts = np.zeros([len(intervals), subspecies.NUM_SUBSPECIES])
for i, prox_species in enumerate(subspecies.iter_subspecies()):
for j, dist_species in enumerate(subspecies.iter_subspecies()):
counts = combo_count_dict[subspecies.combine(prox_species, dist_species)]
species_counts[:, i] += np.diag(counts)
combo_counts[:, :, i * subspecies.NUM_SUBSPECIES + j] = counts
# compute expected combo frequencies from source frequencies
combo_expectations = np.zeros([len(intervals), len(intervals), subspecies.NUM_SUBSPECIES ** 2])
for i in xrange(subspecies.NUM_SUBSPECIES):
for j in xrange(subspecies.NUM_SUBSPECIES):
combo_expectations[:, :, i * subspecies.NUM_SUBSPECIES + j] += \
np.outer(species_counts[:, i], species_counts[:, j])
# normalize expectations using the actual total frequency
sums = np.sum(combo_counts, axis=2)
old_settings = np.seterr(invalid='ignore') # ignore division by 0 errors for intervals with no assigned origin
for i in xrange(subspecies.NUM_SUBSPECIES ** 2):
combo_expectations[:, :, i] = np.true_divide(combo_expectations[:, :, i], sums)
np.seterr(**old_settings)
combo_expectations = np.nan_to_num(combo_expectations)
# do chi-square test
output = []
for i in xrange(len(intervals)):
# only upper triangle is meaningful
for j in xrange(i + 1, len(intervals)):
nonzero_expectations = np.where(combo_expectations[i, j])
chi_sq, p_value = stats.chisquare(
combo_counts[i, j][nonzero_expectations], combo_expectations[i, j][nonzero_expectations])
output.append([
chi_sq,
p_value,
# proximal interval
intervals[i-1],
intervals[i],
# distal interval
intervals[j],
intervals[j-1]
])
return output
def pairwise_frequencies(self, strain_names):
""" For every locus pair and every label pair, count the number of strains which have those
labels at those pairs of loci.
:param strain_names: list of strain names to analyze (must be a subset of the output from preprocess())
"""
output = [[[], [], [], []] for _ in xrange(subspecies.NUM_SUBSPECIES**2)]
for strain_name in strain_names:
intervals, sources = self.sample_dict[strain_name]
for i in xrange(len(intervals)):
# only upper triangle is meaningful
if subspecies.is_known(sources[i]):
for j in xrange(i, len(intervals)):
if subspecies.is_known(sources[j]):
combo_output = output[subspecies.to_ordinal(subspecies.combine(sources[i], sources[j]))]
combo_output[0].append(intervals[i-1])
combo_output[1].append(intervals[i])
combo_output[2].append(intervals[j-1])
combo_output[3].append(intervals[j])
return output, [subspecies.to_color(i, True) for i in xrange(subspecies.NUM_SUBSPECIES**2)]
