def pls_balances_cmd(table_file, metadata_file, category, output_file):
metadata = pd.read_table(metadata_file, index_col=0)
table = load_table(table_file)
table = pd.DataFrame(np.array(table.matrix_data.todense()).T,
index=table.ids(axis='sample'),
columns=table.ids(axis='observation'))
ctable = pd.DataFrame(clr(centralize(table + 1)),
index=table.index,
columns=table.columns)
rfc = PLSRegression(n_components=1)
if metadata[category].dtype != np.float:
cats = np.unique(metadata[category])
groups = (metadata[category] == cats[0]).astype(np.int)
else:
groups = metadata[category]
rfc.fit(X=ctable.values, Y=groups)
pls_df = pd.DataFrame(rfc.x_weights_,
index=ctable.columns,
columns=['PLS1'])
l, r = round_balance(pls_df.values,
means_init=[[pls_df.PLS1.min()], [0],
[pls_df.PLS1.max()]],
n_init=100)
num = pls_df.loc[pls_df.PLS1 > r]
denom = pls_df.loc[pls_df.PLS1 < l]
diff_features = list(num.index.values)
diff_features += list(denom.index.values)
with open(output_file, 'w') as f:
f.write(','.join(diff_features))
def test_biplot(self):
exp = clr(centralize(clr_inv(self.beta)))
res = regression_biplot(self.beta)
self.assertIsInstance(res, OrdinationResults)
u = res.samples.values
v = res.features.values.T
npt.assert_allclose(u @ v, np.array(exp), atol=0.5, rtol=0.5)
def dendrogram_heatmap(output_dir: str, table: pd.DataFrame,
tree: TreeNode, metadata: MetadataCategory,
ndim=10, method='clr', color_map='viridis'):
nodes = [n.name for n in tree.levelorder() if not n.is_tip()]
nlen = min(ndim, len(nodes))
numerator_color, denominator_color = '#fb9a99', '#e31a1c'
highlights = pd.DataFrame([[numerator_color, denominator_color]] * nlen,
index=nodes[:nlen])
if method == 'clr':
mat = pd.DataFrame(clr(centralize(table)),
index=table.index,
columns=table.columns)
elif method == 'log':
mat = pd.DataFrame(np.log(table),
index=table.index,
columns=table.columns)
# TODO: There are a few hard-coded constants here
# will need to have some adaptive defaults set in the future
fig = heatmap(mat, tree, metadata.to_series(), highlights, cmap=color_map,
highlight_width=0.01, figsize=(12, 8))
fig.savefig(os.path.join(output_dir, 'heatmap.svg'))
fig.savefig(os.path.join(output_dir, 'heatmap.pdf'))
css = r"""
.square {
float: left;
width: 100px;
height: 20px;
margin: 5px;
border: 1px solid rgba(0, 0, 0, .2);
}
.numerator {
background: %s;
}
.denominator {
background: %s;
}
""" % (numerator_color, denominator_color)
index_fp = os.path.join(output_dir, 'index.html')
with open(index_fp, 'w') as index_f:
index_f.write('<html><body>\n')
index_f.write('<h1>Dendrogram heatmap</h1>\n')
index_f.write('<img src="heatmap.svg" alt="heatmap">')
index_f.write('<a href="heatmap.pdf">')
index_f.write('Download as PDF</a><br>\n')
index_f.write('<style>%s</style>' % css)
index_f.write('<div class="square numerator">'
'Numerator<br/></div>')
index_f.write('<div class="square denominator">'
'Denominator<br/></div>')
index_f.write('</body></html>\n')
def balance_classify(table, cats, num_folds, **init_kwds):
"""
Builds a balance classifier. If categorical, it is assumed
that the classes are binary.
"""
skf = KFold(n_splits=num_folds, shuffle=True)
ctable = pd.DataFrame(clr(centralize(table)),
index=table.index,
columns=table.columns)
cv = pd.DataFrame(columns=['Q2', 'AUROC'], index=np.arange(num_folds))
for i, (train, test) in enumerate(skf.split(ctable.values, cats.values)):
X_train, X_test = ctable.iloc[train], ctable.iloc[test]
Y_train, Y_test = cats.iloc[train], cats.iloc[test]
plsc = PLSRegression(n_components=1)
plsc.fit(X=X_train, Y=Y_train)
pls_df = pd.DataFrame(plsc.x_weights_,
index=ctable.columns,
columns=['PLS1'])
l, r = round_balance(pls_df, **init_kwds)
denom = pls_df.loc[pls_df.PLS1 < l]
num = pls_df.loc[pls_df.PLS1 > r]
# make the prediction and evaluate the accuracy
idx = table.index[test]
pls_balance = (np.log(table.loc[idx, num.index] + 1).mean(axis=1) -
np.log(table.loc[idx, denom.index] + 1).mean(axis=1))
group_fpr, group_tpr, thresholds = roc_curve(y_true=1 -
(Y_test == 1).astype(int),
y_score=pls_balance)
auroc = auc(group_tpr, group_fpr)
press = ((pls_balance - Y_test)**2).sum()
tss = ((Y_test.mean() - Y_test)**2).sum()
Q2 = 1 - (press / tss)
cv.loc[i, 'Q2'] = Q2
cv.loc[i, 'AUROC'] = auroc
# build model on entire dataset
plsc = PLSRegression(n_components=1)
plsc.fit(X=table.values, Y=cats.values)
pls_df = pd.DataFrame(plsc.x_weights_,
index=ctable.columns,
columns=['PLS1'])
l, r = round_balance(pls_df, **init_kwds)
denom = pls_df.loc[pls_df.PLS1 < l]
num = pls_df.loc[pls_df.PLS1 > r]
pls_balance = (np.log(table.loc[:, num.index]).mean(axis=1) -
np.log(table.loc[:, denom.index]).mean(axis=1))
return num, denom, pls_balance, cv
def balance_regression(table, cats, num_folds, **init_kwds):
"""
Builds a balance classifier. If categorical, it is assumed
that the classes are binary.
"""
skf = KFold(n_splits=num_folds, shuffle=True)
cats = cats * -1 # wtf??
ctable = pd.DataFrame(clr(centralize(table)),
index=table.index,
columns=table.columns)
cv = pd.DataFrame(columns=['Q2'], index=np.arange(num_folds))
for i, (train, test) in enumerate(skf.split(ctable.values, cats.values)):
X_train, X_test = ctable.iloc[train], ctable.iloc[test]
Y_train, Y_test = cats.iloc[train], cats.iloc[test]
plsc = PLSRegression(n_components=1)
plsc.fit(X=X_train, Y=Y_train)
pls_df = pd.DataFrame(plsc.x_weights_,
index=ctable.columns,
columns=['PLS1'])
l, r = round_balance(pls_df, **init_kwds)
denom = pls_df.loc[pls_df.PLS1 < l]
num = pls_df.loc[pls_df.PLS1 > r]
idx = table.index[train]
pls_balance = (np.log(table.loc[idx, num.index] + 1).mean(axis=1) -
np.log(table.loc[idx, denom.index] + 1).mean(axis=1))
b_, int_, _, _, _ = linregress(pls_balance, Y_train)
idx = table.index[test]
pls_balance = (np.log(table.loc[idx, num.index] + 1).mean(axis=1) -
np.log(table.loc[idx, denom.index] + 1).mean(axis=1))
pred = pls_balance * b_ + int_
press = ((pred - Y_test)**2).sum()
tss = ((Y_test.mean() - Y_test)**2).sum()
Q2 = 1 - (press / tss)
cv.loc[i, 'Q2'] = Q2
# build model on entire dataset
plsc = PLSRegression(n_components=1)
plsc.fit(X=table.values, Y=cats.values)
pls_df = pd.DataFrame(plsc.x_weights_,
index=ctable.columns,
columns=['PLS1'])
l, r = round_balance(pls_df, **init_kwds)
denom = pls_df.loc[pls_df.PLS1 < l]
num = pls_df.loc[pls_df.PLS1 > r]
pls_balance = (np.log(table.loc[:, num.index]).mean(axis=1) -
np.log(table.loc[:, denom.index]).mean(axis=1))
return num, denom, pls_balance, cv
def test_center_log(self):
dat = np.array([[10, 20, 1, 20, 5, 100, 844, 100],
[10, 20, 2, 19, 0, 100, 849, 200],
[10, 20, 3, 18, 5, 100, 844, 300],
[10, 20, 4, 17, 0, 100, 849, 400],
[10, 20, 5, 16, 4, 100, 845, 500],
[10, 20, 6, 15, 0, 100, 849, 600],
[10, 20, 7, 14, 3, 100, 846, 700],
[10, 20, 8, 13, 0, 100, 849, 800],
[10, 20, 9, 12, 7, 100, 842, 900]]) + 1
obs = self.test2.center_log()
exp = clr(dat)
assert_array_almost_equal(exp, obs.data)
obs = self.test2.center_log(centralize=True)
exp = clr(centralize(dat))
assert_array_almost_equal(exp, obs.data)
def regression_biplot(coefficients: pd.DataFrame) -> skbio.OrdinationResults:
coefs = clr(centralize(clr_inv(coefficients)))
u, s, v = np.linalg.svd(coefs)
pc_ids = ['PC%d' % i for i in range(len(s))]
samples = pd.DataFrame(u[:, :len(s)] @ np.diag(s),
columns=pc_ids, index=coefficients.index)
features = pd.DataFrame(v.T[:, :len(s)],
columns=pc_ids, index=coefficients.columns)
short_method_name = 'regression_biplot'
long_method_name = 'Multinomial regression biplot'
eigvals = pd.Series(s, index=pc_ids)
proportion_explained = eigvals / eigvals.sum()
res = OrdinationResults(short_method_name, long_method_name, eigvals,
samples=samples, features=features,
proportion_explained=proportion_explained)
return res
def test_centralize(self):
cmat = centralize(closure(self.data1))
npt.assert_allclose(cmat,
np.array([[0.22474487, 0.22474487, 0.55051026],
[0.41523958, 0.41523958, 0.16952085]]))
cmat = centralize(closure(self.data5))
npt.assert_allclose(cmat,
np.array([[0.22474487, 0.22474487, 0.55051026],
[0.41523958, 0.41523958, 0.16952085]]))
with self.assertRaises(ValueError):
centralize(self.bad1)
with self.assertRaises(ValueError):
centralize(self.bad2)
centralize(self.data1)
npt.assert_allclose(self.data1,
np.array([[2, 2, 6],
[4, 4, 2]]))
def test_center_log_ration(self):
from skbio.stats.composition import clr, centralize
dat = np.array([[10, 20, 1, 20, 5, 100, 844, 100],
[10, 20, 2, 19, 0, 100, 849, 200],
[10, 20, 3, 18, 5, 100, 844, 300],
[10, 20, 4, 17, 0, 100, 849, 400],
[10, 20, 5, 16, 4, 100, 845, 500],
[10, 20, 6, 15, 0, 100, 849, 600],
[10, 20, 7, 14, 3, 100, 846, 700],
[10, 20, 8, 13, 0, 100, 849, 800],
[10, 20, 9, 12, 7, 100, 842, 900]]) + 1
obs = self.test2.center_log_ratio()
exp = clr(dat)
assert_array_almost_equal(exp, obs.data)
obs = self.test2.center_log_ratio(centralize=True)
exp = clr(centralize(dat))
assert_array_almost_equal(exp, obs.data)
def test_centralize(self):
cmat = centralize(closure(self.cdata1))
npt.assert_allclose(
cmat,
np.array([[0.22474487, 0.22474487, 0.55051026],
[0.41523958, 0.41523958, 0.16952085]]))
cmat = centralize(closure(self.cdata5))
npt.assert_allclose(
cmat,
np.array([[0.22474487, 0.22474487, 0.55051026],
[0.41523958, 0.41523958, 0.16952085]]))
with self.assertRaises(ValueError):
centralize(self.bad1)
with self.assertRaises(ValueError):
centralize(self.bad2)
# make sure that inplace modification is not occurring
centralize(self.cdata1)
npt.assert_allclose(self.cdata1, np.array([[2, 2, 6], [4, 4, 2]]))
def test_centralize(self):
cmat = centralize(closure(self.cdata1))
npt.assert_allclose(cmat,
np.array([[0.22474487, 0.22474487, 0.55051026],
[0.41523958, 0.41523958, 0.16952085]]))
cmat = centralize(closure(self.cdata5))
npt.assert_allclose(cmat,
np.array([[0.22474487, 0.22474487, 0.55051026],
[0.41523958, 0.41523958, 0.16952085]]))
with self.assertRaises(ValueError):
centralize(self.bad1)
with self.assertRaises(ValueError):
centralize(self.bad2)
# make sure that inplace modification is not occurring
centralize(self.cdata1)
npt.assert_allclose(self.cdata1,
np.array([[2, 2, 6],
[4, 4, 2]]))
def dendrogram_heatmap(output_dir: str,
table: pd.DataFrame,
tree: TreeNode,
metadata: MetadataCategory,
ndim=10,
method='clr',
color_map='viridis'):
nodes = [n.name for n in tree.levelorder()]
nlen = min(ndim, len(nodes))
highlights = pd.DataFrame([['#00FF00', '#FF0000']] * nlen,
index=nodes[:nlen])
if method == 'clr':
mat = pd.DataFrame(clr(centralize(table)),
index=table.index,
columns=table.columns)
elif method == 'log':
mat = pd.DataFrame(np.log(table),
index=table.index,
columns=table.columns)
# TODO: There are a few hard-coded constants here
# will need to have some adaptive defaults set in the future
fig = heatmap(mat,
tree,
metadata.to_series(),
highlights,
cmap=color_map,
highlight_width=0.01,
figsize=(12, 8))
fig.savefig(os.path.join(output_dir, 'heatmap.svg'))
index_fp = os.path.join(output_dir, 'index.html')
with open(index_fp, 'w') as index_f:
index_f.write('<html><body>\n')
index_f.write('<h1>Dendrogram heatmap</h1>\n')
index_f.write('<img src="heatmap.svg" alt="heatmap">')
index_f.write('</body></html>\n')
def balance_taxonomy(output_dir: str, table: pd.DataFrame, tree: TreeNode,
taxonomy: pd.DataFrame,
balance_name: str,
pseudocount: float = 0.5,
taxa_level: int = 0,
n_features: int = 10,
threshold: float = None,
metadata: qiime2.MetadataColumn = None) -> None:
if threshold is not None and isinstance(metadata,
qiime2.CategoricalMetadataColumn):
raise ValueError('Categorical metadata column detected. Only specify '
'a threshold when using a numerical metadata column.')
# make sure that the table and tree match up
table, tree = match_tips(add_pseudocount(table, pseudocount), tree)
# parse out headers for taxonomy
taxa_data = list(taxonomy['Taxon'].apply(lambda x: x.split(';')).values)
taxa_df = pd.DataFrame(taxa_data, index=taxonomy.index)
# fill in NAs
def f(x):
y = np.array(list(map(lambda k: k is not None, x)))
i = max(0, np.where(y)[0][-1])
x[np.logical_not(y)] = [x[i]] * np.sum(np.logical_not(y))
return x
taxa_df = taxa_df.apply(f, axis=1)
num_clade = tree.find(balance_name).children[NUMERATOR]
denom_clade = tree.find(balance_name).children[DENOMINATOR]
if num_clade.is_tip():
num_features = pd.DataFrame(
{num_clade.name: taxa_df.loc[num_clade.name]}
).T
r = 1
else:
num_features = taxa_df.loc[num_clade.subset()]
r = len(list(num_clade.tips()))
if denom_clade.is_tip():
denom_features = pd.DataFrame(
{denom_clade.name: taxa_df.loc[denom_clade.name]}
).T
s = 1
else:
denom_features = taxa_df.loc[denom_clade.subset()]
s = len(list(denom_clade.tips()))
b = (np.log(table.loc[:, num_features.index]).mean(axis=1) -
np.log(table.loc[:, denom_features.index]).mean(axis=1))
b = b * np.sqrt(r * s / (r + s))
balances = pd.DataFrame(b, index=table.index,
columns=[balance_name])
# the actual colors for the numerator and denominator
num_color = sns.color_palette("Paired")[0]
denom_color = sns.color_palette("Paired")[1]
fig, (ax_num, ax_denom) = plt.subplots(2)
balance_barplots(tree, balance_name, taxa_level, taxa_df,
denom_color=denom_color, num_color=num_color,
axes=(ax_num, ax_denom))
ax_num.set_title(
r'$%s_{numerator} \; taxa \; (%d \; taxa)$' % (
balance_name, len(num_features)))
ax_denom.set_title(
r'$%s_{denominator} \; taxa \; (%d \; taxa)$' % (
balance_name, len(denom_features)))
ax_denom.set_xlabel('Number of unique taxa')
plt.tight_layout()
fig.savefig(os.path.join(output_dir, 'barplots.svg'))
fig.savefig(os.path.join(output_dir, 'barplots.pdf'))
dcat = None
multiple_cats = False
if metadata is not None:
fig2, ax = plt.subplots()
c = metadata.to_series()
data, c = match(balances, c)
data[c.name] = c
y = data[balance_name]
# check if continuous
if isinstance(metadata, qiime2.NumericMetadataColumn):
ax.scatter(c.values, y)
ax.set_xlabel(c.name)
if threshold is None:
threshold = c.mean()
dcat = c.apply(
lambda x: '%s < %f' % (c.name, threshold)
if x < threshold
else '%s > %f' % (c.name, threshold)
)
sample_palette = pd.Series(sns.color_palette("Set2", 2),
index=dcat.value_counts().index)
elif isinstance(metadata, qiime2.CategoricalMetadataColumn):
sample_palette = pd.Series(
sns.color_palette("Set2", len(c.value_counts())),
index=c.value_counts().index)
try:
pd.to_numeric(metadata.to_series())
except ValueError:
pass
else:
raise ValueError('Categorical metadata column '
f'{metadata.name!r} contains only numerical '
'values. At least one value must be '
'non-numerical.')
balance_boxplot(balance_name, data, y=c.name, ax=ax,
palette=sample_palette)
if len(c.value_counts()) > 2:
warnings.warn(
'More than 2 categories detected in categorical metadata '
'column. Proportion plots will not be displayed',
stacklevel=2)
multiple_cats = True
else:
dcat = c
else:
# Some other type of MetadataColumn
raise NotImplementedError()
ylabel = (r"$%s = \ln \frac{%s_{numerator}}"
"{%s_{denominator}}$") % (balance_name,
balance_name,
balance_name)
ax.set_title(ylabel, rotation=0)
ax.set_ylabel('log ratio')
fig2.savefig(os.path.join(output_dir, 'balance_metadata.svg'))
fig2.savefig(os.path.join(output_dir, 'balance_metadata.pdf'))
if not multiple_cats:
# Proportion plots
# first sort by clr values and calculate average fold change
ctable = pd.DataFrame(clr(centralize(table)),
index=table.index, columns=table.columns)
left_group = dcat.value_counts().index[0]
right_group = dcat.value_counts().index[1]
lidx, ridx = (dcat == left_group), (dcat == right_group)
if b.loc[lidx].mean() > b.loc[ridx].mean():
# double check ordering and switch if necessary
# careful - the left group is also commonly associated with
# the denominator.
left_group = dcat.value_counts().index[1]
right_group = dcat.value_counts().index[0]
lidx, ridx = (dcat == left_group), (dcat == right_group)
# we are not performing a statistical test here
# we're just trying to figure out a way to sort the data.
num_fold_change = ctable.loc[:, num_features.index].apply(
lambda x: ttest_ind(x[ridx], x[lidx])[0])
num_fold_change = num_fold_change.sort_values(
ascending=False
)
denom_fold_change = ctable.loc[:, denom_features.index].apply(
lambda x: ttest_ind(x[ridx], x[lidx])[0])
denom_fold_change = denom_fold_change.sort_values(
ascending=True
)
metadata = pd.DataFrame({dcat.name: dcat})
top_num_features = num_fold_change.index[:n_features]
top_denom_features = denom_fold_change.index[:n_features]
fig3, (ax_denom, ax_num) = plt.subplots(1, 2)
proportion_plot(
table, metadata,
category=metadata.columns[0],
left_group=left_group,
right_group=right_group,
feature_metadata=taxa_df,
label_col=taxa_level,
num_features=top_num_features,
denom_features=top_denom_features,
# Note that the syntax is funky and counter
# intuitive. This will need to be properly
# fixed here
# https://github.com/biocore/gneiss/issues/244
num_color=sample_palette.loc[right_group],
denom_color=sample_palette.loc[left_group],
axes=(ax_num, ax_denom))
# The below is overriding the default colors in the
# numerator / denominator this will also need to be fixed in
# https://github.com/biocore/gneiss/issues/244
max_ylim, min_ylim = ax_denom.get_ylim()
num_h, denom_h = n_features, n_features
space = (max_ylim - min_ylim) / (num_h + denom_h)
ymid = (max_ylim - min_ylim) * num_h
ymid = ymid / (num_h + denom_h) - 0.5 * space
ax_denom.axhspan(min_ylim, ymid,
facecolor=num_color,
zorder=0)
ax_denom.axhspan(ymid, max_ylim,
facecolor=denom_color,
zorder=0)
ax_num.axhspan(min_ylim, ymid,
facecolor=num_color,
zorder=0)
ax_num.axhspan(ymid, max_ylim,
facecolor=denom_color,
zorder=0)
fig3.subplots_adjust(
# the left side of the subplots of the figure
left=0.3,
# the right side of the subplots of the figure
right=0.9,
# the bottom of the subplots of the figure
bottom=0.1,
# the top of the subplots of the figure
top=0.9,
# the amount of width reserved for blank space
# between subplots
wspace=0,
# the amount of height reserved for white space
# between subplots
hspace=0.2,
)
fig3.savefig(os.path.join(output_dir, 'proportion_plot.svg'))
fig3.savefig(os.path.join(output_dir, 'proportion_plot.pdf'))
index_fp = os.path.join(output_dir, 'index.html')
with open(index_fp, 'w') as index_f:
index_f.write('<html><body>\n')
if metadata is not None:
index_f.write('<h1>Balance vs %s </h1>\n' % c.name)
index_f.write(('<img src="balance_metadata.svg" '
'alt="barplots">\n\n'
'<a href="balance_metadata.pdf">'
'Download as PDF</a><br>\n'))
if not multiple_cats:
index_f.write('<h1>Proportion Plot </h1>\n')
index_f.write(('<img src="proportion_plot.svg" '
'alt="proportions">\n\n'
'<a href="proportion_plot.pdf">'
'Download as PDF</a><br>\n'))
index_f.write(('<h1>Balance Taxonomy</h1>\n'
'<img src="barplots.svg" alt="barplots">\n\n'
'<a href="barplots.pdf">'
'Download as PDF</a><br>\n'
'<h3>Numerator taxa</h3>\n'
'<a href="numerator.csv">\n'
'Download as CSV</a><br>\n'
'<h3>Denominator taxa</h3>\n'
'<a href="denominator.csv">\n'
'Download as CSV</a><br>\n'))
num_features.to_csv(os.path.join(output_dir, 'numerator.csv'),
header=True, index=True)
denom_features.to_csv(os.path.join(output_dir, 'denominator.csv'),
header=True, index=True)
index_f.write('</body></html>\n')
microbe_iv['group'] = microbe_iv['group'].map(catdict)
metabolite_iv['group'] = metabolite_iv['group'].map(catdict)
# highlight features with p-value <= 0.001
max_pval = 0.001
microbe_iv.loc[microbe_iv.pval > max_pval, 'group'] = 'None'
print('Number of significant microbes: %d' %
microbe_iv[microbe_iv['group'] != 'None'].shape[0])
metabolite_iv.loc[metabolite_iv.pval > max_pval, 'group'] = 'None'
print('Number of significant metabolites: %d' %
metabolite_iv[metabolite_iv['group'] != 'None'].shape[0])
plssvd = PLSSVD(n_components=3)
plssvd.fit(X=clr(centralize(multiplicative_replacement(microbes))),
Y=clr(centralize(multiplicative_replacement(metabolites))))
def standardize(A):
A = (A - np.mean(A, axis=0)) / np.std(A, axis=0)
return A
pls_microbes = pd.DataFrame(standardize(plssvd.x_weights_),
columns=['PCA1', 'PCA2', 'PCA3'],
index=microbes.columns)
pls_metabolites = pd.DataFrame(standardize(plssvd.y_weights_),
columns=['PCA1', 'PCA2', 'PCA3'],
index=metabolites.columns)