def rpca(in_biom: str, output_dir: str,
min_sample_depth: int, rank: int) -> None:
""" Runs RPCA with an rclr preprocessing step"""
# import table
table = load_table(in_biom)
# filter sample to min depth
def sample_filter(val, id_, md): return sum(val) > min_sample_depth
table = table.filter(sample_filter, axis='sample')
table = table.to_dataframe().T.drop_duplicates()
# rclr for saving the transformed OTU table (RSC edited)
tablefit = rclr().fit_transform(table.copy())
U,s,V = OptSpace().fit_transform(tablefit)
tablefit = np.dot(np.dot(U, s), V.T)
tablefit = pd.DataFrame(tablefit.T, index=table.columns, columns=table.index)
with open(os.path.join(output_dir, 'rclr_OTUtable.txt'), 'w'):
tablefit.to_csv(os.path.join(output_dir, 'rclr_OTUtable.txt'), sep='\t', index_label='OTU_ID')
# rclr preprocessing and OptSpace (RPCA)
opt = OptSpace(rank=rank).fit(rclr().fit_transform(table.copy()))
rename_cols = {i - 1: 'PC' + str(i) for i in range(1, rank + 1)}
# Feature Loadings
feature_loading = pd.DataFrame(opt.feature_weights, index=table.columns)
feature_loading = feature_loading.rename(columns=rename_cols)
feature_loading.sort_values('PC1', inplace=True, ascending=True)
# Sample Loadings
sample_loading = pd.DataFrame(opt.sample_weights, index=table.index)
sample_loading = sample_loading.rename(columns=rename_cols)
proportion_explained = pd.Series(opt.explained_variance_ratio,
index=list(rename_cols.values()))
eigvals = pd.Series(opt.eigenvalues,
index=list(rename_cols.values()))
# save ordination results
ord_res = OrdinationResults(
'PCoA',
'Principal Coordinate Analysis',
eigvals.copy(),
sample_loading.copy(),
features=feature_loading.copy(),
proportion_explained=proportion_explained.copy())
# write files to output folder
ord_res.write(os.path.join(output_dir, 'RPCA_Ordination.txt'))
# save distance matrix
dist_res = skbio.stats.distance.DistanceMatrix(
opt.distance, ids=sample_loading.index)
dist_res.write(os.path.join(output_dir, 'RPCA_distance.txt'))
return
def rpca(
table: biom.Table,
rank: int = 3,
min_sample_count: int = 500,
min_feature_count: int = 10,
iterations: int = 5
) -> (skbio.OrdinationResults, skbio.DistanceMatrix):
""" Runs RPCA with an rclr preprocessing step"""
# filter sample to min depth
def sample_filter(val, id_, md):
return sum(val) > min_sample_count
table = table.filter(sample_filter, axis='sample')
table = table.to_dataframe().T.drop_duplicates()
table = table.T[table.sum() > min_feature_count].T
# rclr preprocessing and OptSpace (RPCA)
opt = OptSpace(rank=rank, iteration=iterations).fit(rclr().fit_transform(
table.copy()))
rename_cols = {i - 1: 'PC' + str(i) for i in range(1, rank + 1)}
# Feature Loadings
feature_loading = pd.DataFrame(opt.feature_weights, index=table.columns)
feature_loading = feature_loading.rename(columns=rename_cols)
feature_loading.sort_values('PC1', inplace=True, ascending=True)
# Sample Loadings
sample_loading = pd.DataFrame(opt.sample_weights, index=table.index)
sample_loading = sample_loading.rename(columns=rename_cols)
# % var explained
proportion_explained = pd.Series(opt.explained_variance_ratio,
index=list(rename_cols.values()))
# eigan-vals
eigvals = pd.Series(opt.eigenvalues, index=list(rename_cols.values()))
# if the rank is two add PC3 of zeros
if rank == 2:
feature_loading['PC3'] = [0] * len(feature_loading.index)
sample_loading['PC3'] = [0] * len(sample_loading.index)
eigvals.loc['PC3'] = 0
proportion_explained.loc['PC3'] = 0
# save ordination results
short_method_name = 'rpca_biplot'
long_method_name = '(Robust Aitchison) RPCA Biplot'
ord_res = skbio.OrdinationResults(
short_method_name,
long_method_name,
eigvals.copy(),
samples=sample_loading.copy(),
features=feature_loading.copy(),
proportion_explained=proportion_explained.copy())
# save distance matrix
dist_res = skbio.stats.distance.DistanceMatrix(opt.distance,
ids=sample_loading.index)
return ord_res, dist_res
def tensor_rclr(T):
# flatten, transform, and reshape
T_rclr = np.concatenate([T[i, :, :].T for i in range(T.shape[0])], axis=0)
T_rclr = rclr().fit_transform(T_rclr)
T_rclr = np.dstack([
T_rclr[(i - 1) * T.shape[-1]:(i) * T.shape[-1]]
for i in range(1, T.shape[0] + 1)
])
T_rclr[np.isnan(T_rclr)] = 0
return T_rclr
def setUp(self):
self.cdata1 = np.array([[2, 2, 6], [4, 4, 2]])
self.cdata2 = [[3, 3, 0], [0, 4, 2]]
self.true2 = np.array([[0.0, 0.0, np.nan],
[np.nan, 0.34657359, -0.34657359]])
self.bad1 = np.array([1, 2, -1])
self.bad1
self._rclr = rclr()
self._inv = inverse_rclr()
pass
def rpca(adata):
from deicode.matrix_completion import MatrixCompletion
from deicode.preprocessing import rclr
min_samples = max(3, np.floor(n_samples * 0.1))
sc.pp.filter_genes(adata, min_cells=min_samples)
X = rclr(adata.raw.X)
opt = MatrixCompletion(n_components=n_comps, max_iterations=10).fit(X)
n_components = opt.s.shape[0]
X = opt.sample_weights @ opt.s @ opt.feature_weights.T
X = X - X.mean(axis=0)
X = X - X.mean(axis=1).reshape(-1, 1)
adata.obsm['X_deicode'] = sc.tl.pca(X,
svd_solver='arpack',
n_comps=n_comps)
return adata
def create_test_table():
_, test_table = build_block_model(rank=2,
hoced=20,
hsced=20,
spar=2e3,
C_=2e3,
num_samples=50,
num_features=500,
mapping_on=False)
# the rclr is tested in other places
# this is just used as input into
# the OptSpace tests
test_table = np.array(test_table)
table_rclr = rclr(test_table)
return test_table, table_rclr
def get_distance_matrix(otutabledf, index, method, rank, logger_ins, training_index):
if method == "deicode":
logger_ins.info("method is deicode, and the rank is", rank)
rclr_obj = rclr()
table_norm = rclr_obj.fit_transform(copy.deepcopy(otutabledf))
opt = OptSpace(rank=rank,iteration=10,tol=1e-5).fit(table_norm)
sample_loading = pd.DataFrame(opt.sample_weights, index=index)
rapca = DistanceMatrix(distance.cdist(sample_loading.values, sample_loading.values, 'euclidean'))
else:
table_norm = np.asarray(otutabledf / np.sum(otutabledf, 1))
logger_ins.info("Using the braycurtis way of calculating distances between samples")
logger_ins.info("the shape of the normalized table is", table_norm.shape)
logger_ins.info("The sum of columns of the table is", set(list(table_norm.sum(1).squeeze())), "and the size is", table_norm.sum(1).shape)
sample_loading = pd.DataFrame(table_norm, index=index)
rapca = DistanceMatrix(pdist(sample_loading.values, 'braycurtis'))
return rapca
def test_rclr_nan_raises(self):
"""Test rclr ValueError on missing (as nan)."""
# test nan throw value error
with self.assertRaises(ValueError):
rclr(self.bad3)
def test_rclr_inf_raises(self):
"""Test rclr ValueError on undefined."""
# test undefined throw value error
with self.assertRaises(ValueError):
rclr(self.bad2)
def test_rclr_negative_raises(self):
"""Test rclr ValueError on negative."""
# test negatives throw value error
with self.assertRaises(ValueError):
rclr(self.bad1)
def test_rclr_sparse(self):
"""Test rclr on sparse data."""
# test a case with zeros :)
cmat = rclr(self.cdata2)
npt.assert_allclose(cmat, self.true2)
def test_rclr_dense(self):
"""Test rclr and clr are the same on dense datasets."""
# test clr works the same if there are no zeros
cmat = rclr(self.cdata1)
npt.assert_allclose(cmat, clr(self.cdata1.copy()))
def rpca(
table: biom.Table,
n_components: int = DEFAULT_RANK,
min_sample_count: int = DEFAULT_MSC,
min_feature_count: int = DEFAULT_MFC,
max_iterations: int = DEFAULT_ITERATIONS
) -> (skbio.OrdinationResults, skbio.DistanceMatrix):
"""Runs RPCA with an rclr preprocessing step.
This code will be run by both the standalone and QIIME 2 versions of
DEICODE.
"""
# filter sample to min depth
def sample_filter(val, id_, md):
return sum(val) > min_sample_count
def observation_filter(val, id_, md):
return sum(val) > min_feature_count
# filter and import table
table = table.filter(observation_filter, axis='observation')
table = table.filter(sample_filter, axis='sample')
table = table.to_dataframe().T
if len(table.index) != len(set(table.index)):
raise ValueError('Data-table contains duplicate indices')
if len(table.columns) != len(set(table.columns)):
raise ValueError('Data-table contains duplicate columns')
# rclr preprocessing and OptSpace (RPCA)
opt = MatrixCompletion(n_components=n_components,
max_iterations=max_iterations).fit(rclr(table))
rename_cols = ['PC' + str(i + 1) for i in range(n_components)]
X = opt.sample_weights @ opt.s @ opt.feature_weights.T
X = X - X.mean(axis=0)
X = X - X.mean(axis=1).reshape(-1, 1)
u, s, v = svd(X)
u = u[:, :n_components]
v = v.T[:, :n_components]
p = s**2 / np.sum(s**2)
p = p[:n_components]
s = s[:n_components]
feature_loading = pd.DataFrame(v, index=table.columns, columns=rename_cols)
sample_loading = pd.DataFrame(u, index=table.index, columns=rename_cols)
# % var explained
proportion_explained = pd.Series(p, index=rename_cols)
# get eigenvalues
eigvals = pd.Series(s, index=rename_cols)
# if the n_components is two add PC3 of zeros
# this is referenced as in issue in
# <https://github.com/biocore/emperor/commit
# /a93f029548c421cb0ba365b4294f7a5a6b0209ce>
# discussed in DEICODE -- PR#29
if n_components == 2:
feature_loading['PC3'] = [0] * len(feature_loading.index)
sample_loading['PC3'] = [0] * len(sample_loading.index)
eigvals.loc['PC3'] = 0
proportion_explained.loc['PC3'] = 0
# save ordination results
short_method_name = 'rpca_biplot'
long_method_name = '(Robust Aitchison) RPCA Biplot'
ord_res = skbio.OrdinationResults(
short_method_name,
long_method_name,
eigvals.copy(),
samples=sample_loading.copy(),
features=feature_loading.copy(),
proportion_explained=proportion_explained.copy())
# save distance matrix
dist_res = skbio.stats.distance.DistanceMatrix(opt.distance,
ids=sample_loading.index)
return ord_res, dist_res
), :].copy().T
#get the base truth data for that subset
basetmp_sub = base_truth.loc[(
rank_,
power_,
depth_,
), :].copy().T
# sub sampled
subtmp_sub = subtmp.copy()
#meta on cluster
meta = np.array([1] * int(subtmp.shape[0] / 2) +
[2] * int(subtmp.shape[0] / 2)).T
meta = pd.DataFrame(meta, index=subtmp.index, columns=['group'])
# test KL with rcl
X_sparse = rclr().fit_transform(subtmp_sub.copy())
U, s, V = OptSpace(iteration=1000).fit_transform(X_sparse)
clr_res = clr_inv(np.dot(np.dot(U, s), V.T))
# use just kl_div here because already closed
kl_clr = entropy(closure(basetmp_sub).T, clr_res.T).mean()
results[(rank_, power_, depth_, 'rclr', 'KL-Div')] = [kl_clr]
# test KL without rclr
X_spn = np.array(subtmp_sub.copy()).astype(float)
X_spn[X_spn == 0] = np.nan
U_, s_, V_ = OptSpace(iteration=1000).fit_transform(X_spn)
res_raw = np.dot(np.dot(U_, s_), V_.T)
res_raw[res_raw <= 0] = 1
kl_raw = entropy(closure(basetmp_sub).T, closure(res_raw).T).mean()
results[(rank_, power_, depth_, 'Raw Counts', 'KL-Div')] = [kl_raw]
control_libs += list(abund_df3.index[abund_df3.index.str.contains("CDSBBR")])
otus_to_strip_c, otus_to_strip_nc = set(), set()
for c in control_libs:
otus_to_strip_nc.update(abund_df3.columns[abund_df3.loc[c, :] > 0])
otus_to_strip_c.update(abund_df5.columns[abund_df5.loc[c, :] > 0])
print("{}/{} are to be removed".format(len(otus_to_strip_nc), len(otus_to_strip_c)))
# remove OTUS in 50 unecessary samples
abund_df_c = abund_df5.loc[:, ~abund_df5.columns.isin(otus_to_strip_c)]
abund_df_c2 = abund_df_c[abund_df_c.sum(1)!=0]
abund_df_nc = abund_df3.loc[:, ~abund_df3.columns.isin(otus_to_strip_nc)]
abund_df_nc2 = abund_df_nc[abund_df_nc.sum(1)!=0]
rename_cols = {i - 1: 'PC' + str(i) for i in range(1, 3)}
rclr_mat = rclr().fit_transform(abund_df_c2.values)
rclr_mat2 = rclr().fit_transform(abund_df_nc2.values)
opt=OptSpace(rank=2).fit(rclr_mat)
# PCA of new mat
feature_loading = pd.DataFrame(opt.feature_weights, index=abund_df_c2.columns).rename(columns=rename_cols)
feature_loading.sort_values('PC1', inplace=True, ascending=True)
sample_loading = pd.DataFrame(opt.sample_weights, index=abund_df_c2.index).rename(columns=rename_cols)
# PCA of old mat
opt2=OptSpace(rank=2).fit(rclr_mat2)
fl_clean = pd.DataFrame(opt2.feature_weights, index=abund_df_nc2.columns).rename(columns=rename_cols)
fl_clean.sort_values('PC1', inplace=True, ascending=True)
sl_clean = pd.DataFrame(opt2.sample_weights, index=abund_df_nc2.index).rename(columns=rename_cols)
sub_meta_c_st = meta_df.loc[abund_df_c2.index, 'station'].tolist()
sub_meta_c_date = meta_df.loc[abund_df_c2.index, 'date'].tolist()
sub_meta_c = meta_df.loc[abund_df_c2.index, 'run'].tolist()
sub_meta_nc = meta_df.loc[abund_df_nc2.index, 'run'].tolist()
def rpca(
table: biom.Table,
n_components: Union[int, str] = DEFAULT_RANK,
min_sample_count: int = DEFAULT_MSC,
min_feature_count: int = DEFAULT_MFC,
min_feature_frequency: float = DEFAULT_MFF,
max_iterations: int = DEFAULT_ITERATIONS
) -> (skbio.OrdinationResults, skbio.DistanceMatrix):
"""Runs RPCA with an rclr preprocessing step.
This code will be run by both the standalone and QIIME 2 versions of
DEICODE.
"""
# get shape of table
n_features, n_samples = table.shape
# filter sample to min seq. depth
def sample_filter(val, id_, md):
return sum(val) > min_sample_count
# filter features to min total counts
def observation_filter(val, id_, md):
return sum(val) > min_feature_count
# filter features by N samples presence
def frequency_filter(val, id_, md):
return (np.sum(val > 0) / n_samples) > (min_feature_frequency / 100)
# filter and import table for each filter above
table = table.filter(observation_filter, axis='observation')
table = table.filter(frequency_filter, axis='observation')
table = table.filter(sample_filter, axis='sample')
table = table.to_dataframe().T
# check the table after filtering
if len(table.index) != len(set(table.index)):
raise ValueError('Data-table contains duplicate indices')
if len(table.columns) != len(set(table.columns)):
raise ValueError('Data-table contains duplicate columns')
# Robust-clt (rclr) preprocessing and OptSpace (RPCA)
opt = MatrixCompletion(n_components=n_components,
max_iterations=max_iterations).fit(rclr(table))
# get new n-comp when applicable
n_components = opt.s.shape[0]
# get PC column labels for the skbio OrdinationResults
rename_cols = ['PC' + str(i + 1) for i in range(n_components)]
# get completed matrix for centering
X = opt.sample_weights @ opt.s @ opt.feature_weights.T
# center again around zero after completion
X = X - X.mean(axis=0)
X = X - X.mean(axis=1).reshape(-1, 1)
# re-factor the data
u, s, v = svd(X)
# only take n-components
u = u[:, :n_components]
v = v.T[:, :n_components]
# calc. the new variance using projection
p = s**2 / np.sum(s**2)
p = p[:n_components]
s = s[:n_components]
# save the loadings
robust_clr = pd.DataFrame(X, index=table.index, columns=table.columns)
feature_loading = pd.DataFrame(v, index=table.columns, columns=rename_cols)
sample_loading = pd.DataFrame(u, index=table.index, columns=rename_cols)
# % var explained
proportion_explained = pd.Series(p, index=rename_cols)
# get eigenvalues
eigvals = pd.Series(s, index=rename_cols)
# if the n_components is two add PC3 of zeros
# this is referenced as in issue in
# <https://github.com/biocore/emperor/commit
# /a93f029548c421cb0ba365b4294f7a5a6b0209ce>
# discussed in DEICODE -- PR#29
if n_components == 2:
feature_loading['PC3'] = [0] * len(feature_loading.index)
sample_loading['PC3'] = [0] * len(sample_loading.index)
eigvals.loc['PC3'] = 0
proportion_explained.loc['PC3'] = 0
# save ordination results
short_method_name = 'rpca_biplot'
long_method_name = '(Robust Aitchison) RPCA Biplot'
ord_res = skbio.OrdinationResults(
short_method_name,
long_method_name,
eigvals.copy(),
samples=sample_loading.copy(),
features=feature_loading.copy(),
proportion_explained=proportion_explained.copy())
# save distance matrix
dist_res = skbio.stats.distance.DistanceMatrix(opt.distance,
ids=sample_loading.index)
return ord_res, dist_res, robust_clr
