def generate_test_bench(self, count_file_path, **kwargs):
preserve_columns = kwargs['preserve_columns']
count_file_path = os.path.abspath(count_file_path)
count_rna = self.data_set.get("RNA")
count_adt = self.data_set.get("ADT")
# Shuffle columns
count_rna, original_columns, column_permutation = \
shuffle_and_rename_columns(count_rna, disabled=preserve_columns)
# Remove zero rows
count_rna = count_rna[np.sum(count_rna, axis=1) > 0].copy()
# Save hidden data
make_sure_dir_exists(settings.STORAGE_DIR)
hidden_data_file_path = os.path.join(settings.STORAGE_DIR,
"%s.hidden.pkl.gz" % self.uid)
dump_gzip_pickle([
count_rna.to_sparse(), original_columns, column_permutation,
count_adt.to_sparse(), self.protein_rna_mapping
], hidden_data_file_path)
log("Benchmark hidden data saved to `%s`" % hidden_data_file_path)
make_sure_dir_exists(os.path.dirname(count_file_path))
write_csv(count_rna, count_file_path)
log("Count file saved to `%s`" % count_file_path)
def generate_test_bench(self, count_file_path, **kwargs):
preserve_columns = kwargs['preserve_columns']
count_file_path = os.path.abspath(count_file_path)
count_matrix, classes = self._load_data()
# Remove zero rows
count_matrix = count_matrix[np.sum(count_matrix, axis=1) > 0].copy()
# Shuffle columns
count_matrix, original_columns, column_permutation = \
shuffle_and_rename_columns(count_matrix, disabled=preserve_columns)
# Save hidden data
make_sure_dir_exists(settings.STORAGE_DIR)
hidden_data_file_path = os.path.join(settings.STORAGE_DIR,
"%s.hidden.pkl.gz" % self.uid)
dump_gzip_pickle([
count_matrix.to_sparse(), classes, original_columns,
column_permutation
], hidden_data_file_path)
log("Benchmark hidden data saved to `%s`" % hidden_data_file_path)
make_sure_dir_exists(os.path.dirname(count_file_path))
write_csv(count_matrix, count_file_path)
log("Count file saved to `%s`" % count_file_path)
def generate_test_bench(self, count_file_path, **kwargs):
n_samples = kwargs['n_samples']
dropout_count = kwargs['dropout_count']
min_expression = kwargs['min_expression']
hvg_frac = kwargs['hvg_frac']
preserve_columns = kwargs['preserve_columns']
count_file_path = os.path.abspath(count_file_path)
data = self._load_data(n_samples)
hvg_indices = self.get_hvg_genes(data, hvg_frac)
# Generate elimination mask
non_zero_locations = []
data_values = data.values
for x in hvg_indices:
for y in range(data.shape[1]):
if data_values[x, y] >= min_expression:
non_zero_locations.append((x, y))
del data_values
mask = np.zeros_like(data)
masked_locations = [
non_zero_locations[index] for index in np.random.choice(
len(non_zero_locations), dropout_count, replace=False)
]
for (x, y) in masked_locations:
mask[x, y] = 1
mask = pd.DataFrame(mask, index=data.index, columns=data.columns)
# Elimination
low_quality_data = data * (1 - mask.values)
is_nonzero = np.sum(low_quality_data, axis=1) > 0
mask = mask[is_nonzero].copy()
data = data[is_nonzero].copy()
low_quality_data = low_quality_data[is_nonzero].copy()
# Shuffle columns
low_quality_data, original_columns, column_permutation = \
shuffle_and_rename_columns(low_quality_data, disabled=preserve_columns)
# Save hidden data
make_sure_dir_exists(settings.STORAGE_DIR)
hidden_data_file_path = os.path.join(settings.STORAGE_DIR,
"%s.hidden.pkl.gz" % self.uid)
dump_gzip_pickle([
data.to_sparse(),
mask.to_sparse(), original_columns, column_permutation
], hidden_data_file_path)
log("Benchmark hidden data saved to `%s`" % hidden_data_file_path)
make_sure_dir_exists(os.path.dirname(count_file_path))
write_csv(low_quality_data, count_file_path)
log("Count file saved to `%s`" % count_file_path)
def generate_test_bench(self, count_file_path, **kwargs):
n_samples = kwargs['n_samples']
read_ratio = kwargs['read_ratio']
replce = kwargs['replace']
preserve_columns = kwargs['preserve_columns']
count_file_path = os.path.abspath(count_file_path)
data = self._load_data(n_samples)
# find cumulative distribution (sum)
data_values = data.astype(int).values
n_all_reads = np.sum(data_values)
data_cumsum = np.reshape(np.cumsum(data_values), data_values.shape)
# Sample from original dataset
new_reads = np.sort(
np.random.choice(n_all_reads,
int(read_ratio * n_all_reads),
replace=replce))
low_quality_data = np.zeros_like(data_values)
read_index = 0
for x in range(data_values.shape[0]):
for y in range(data_values.shape[1]):
while read_index < len(
new_reads) and new_reads[read_index] < data_cumsum[x,
y]:
low_quality_data[x, y] += 1
read_index += 1
# Convert to data frame
low_quality_data = pd.DataFrame(low_quality_data,
index=data.index,
columns=data.columns)
# Shuffle columns
low_quality_data, original_columns, column_permutation = \
shuffle_and_rename_columns(low_quality_data, disabled=preserve_columns)
# Remove zero rows
data = data[np.sum(low_quality_data, axis=1) > 0].copy()
low_quality_data = low_quality_data[
np.sum(low_quality_data, axis=1) > 0].copy()
# Save hidden data
make_sure_dir_exists(settings.STORAGE_DIR)
hidden_data_file_path = os.path.join(settings.STORAGE_DIR,
"%s.hidden.pkl.gz" % self.uid)
dump_gzip_pickle([
data.to_sparse(), read_ratio, original_columns, column_permutation
], hidden_data_file_path)
log("Benchmark hidden data saved to `%s`" % hidden_data_file_path)
make_sure_dir_exists(os.path.dirname(count_file_path))
write_csv(low_quality_data, count_file_path)
log("Count file saved to `%s`" % count_file_path)
def _download_data_set(self):
make_sure_dir_exists(self.DATA_SET_DIR_NAME)
download_file_if_not_exists(self.PBMC_RNA_DATA_URL, self.PBMC_RNA_DATA_FILE_PATH, self.PBMC_RNA_DATA_MD5_SUM)
download_file_if_not_exists(self.PBMC_ADT_DATA_URL, self.PBMC_ADT_DATA_FILE_PATH, self.PBMC_ADT_DATA_MD5_SUM)
download_file_if_not_exists(self.PBMC_TRANSFORMED_ADT_DATA_URL, self.PBMC_TRANSFORMED_ADT_DATA_FILE_PATH,
self.PBMC_TRANSFORMED_ADT_DATA_MD5_SUM)
download_file_if_not_exists(self.CBMC_RNA_DATA_URL, self.CBMC_RNA_DATA_FILE_PATH, self.CBMC_RNA_DATA_MD5_SUM)
download_file_if_not_exists(self.CBMC_ADT_DATA_URL, self.CBMC_ADT_DATA_FILE_PATH, self.CBMC_ADT_DATA_MD5_SUM)
download_file_if_not_exists(self.CBMC_TRANSFORMED_ADT_DATA_URL, self.CBMC_TRANSFORMED_ADT_DATA_FILE_PATH,
self.CBMC_TRANSFORMED_ADT_DATA_MD5_SUM)
download_file_if_not_exists(self.CD8_RNA_DATA_URL, self.CD8_RNA_DATA_FILE_PATH, self.CD8_RNA_DATA_MD5_SUM)
download_file_if_not_exists(self.CD8_ADT_DATA_URL, self.CD8_ADT_DATA_FILE_PATH, self.CD8_ADT_DATA_MD5_SUM)
download_file_if_not_exists(self.CD8_TRANSFORMED_ADT_DATA_URL, self.CD8_TRANSFORMED_ADT_DATA_FILE_PATH,
self.CD8_TRANSFORMED_ADT_DATA_MD5_SUM)
def generate_test_bench(self, count_file_path, **kwargs):
count_file_path = os.path.abspath(count_file_path)
rm_ercc = kwargs['rm_ercc']
rm_mt = kwargs['rm_mt']
rm_lq = kwargs['rm_lq']
preserve_columns = kwargs['preserve_columns']
# Load dataset
data = self._load_and_combine_data()
# Remove some rows and columns
if rm_ercc:
remove_list = [
symbol for symbol in data.index.values
if symbol.startswith("ERCC-")
]
data = data.drop(remove_list)
if rm_mt:
remove_list = [
symbol for symbol in data.index.values
if symbol.startswith("mt-")
]
data = data.drop(remove_list)
if rm_lq:
remove_list = data.columns.values[data.sum(axis=0) < 1e6]
data = data.drop(columns=remove_list)
# Remove empty rows
remove_list = data.index.values[data.sum(axis=1) == 0]
data = data.drop(remove_list)
# Shuffle columns
new_data, original_columns, column_permutation = shuffle_and_rename_columns(
data, disabled=preserve_columns)
# Save hidden data
make_sure_dir_exists(settings.STORAGE_DIR)
hidden_data_file_path = os.path.join(settings.STORAGE_DIR,
"%s.hidden.pkl.gz" % self.uid)
dump_gzip_pickle(
[data.to_sparse(), original_columns, column_permutation],
hidden_data_file_path)
log("Benchmark hidden data saved to `%s`" % hidden_data_file_path)
make_sure_dir_exists(os.path.dirname(count_file_path))
write_csv(new_data, count_file_path)
log("Count file saved to `%s`" % count_file_path)
return None
def evaluate_result(self, processed_count_file_path, result_dir,
visualization, **kwargs):
normalization = kwargs['normalization']
transformation = kwargs['transformation']
clear_cache = kwargs['clear_cache']
make_sure_dir_exists(os.path.join(result_dir, "files"))
info = []
# Load hidden state and data
count_matrix, classes, original_columns, column_permutation = self._load_hidden_state(
)
# Load imputed data
imputed_data = read_table_file(processed_count_file_path)
# Restore column names and order
imputed_data = rearrange_and_rename_columns(imputed_data,
original_columns,
column_permutation)
# Data transformations
if np.sum(imputed_data.values < 0) > 0:
log("Observed some negative values!")
imputed_data[imputed_data < 0] = 0
imputed_data = transformations[transformation](
normalizations[normalization](imputed_data))
# Save class details for future
write_csv(classes, os.path.join(result_dir, "files", "classes.csv"))
# Evaluation
metric_results = dict()
embedded_data_file_path = os.path.join(result_dir, "files",
"embedded_data.pkl.gz")
if os.path.exists(embedded_data_file_path) and not clear_cache:
embedded_data = load_gzip_pickle(embedded_data_file_path)
else:
embedded_data = self._get_embeddings(imputed_data)
dump_gzip_pickle(embedded_data, embedded_data_file_path)
log("Evaluating ...")
for class_label in classes.index.values:
class_names = classes.loc[class_label].values
for embedding_name in embedded_data:
emb, emb_2d = embedded_data[embedding_name]
embedding_slug = embedding_name.replace(" ", "_").lower()
k_means = KMeans(n_clusters=len(set(class_names)))
k_means.fit(emb)
clusters = k_means.predict(emb)
embedding_df = pd.DataFrame(emb)
embedding_df["X"] = emb_2d[:, 0]
embedding_df["Y"] = emb_2d[:, 1]
embedding_df["class"] = class_names
embedding_df["k_means_clusters"] = clusters
write_csv(
embedding_df,
os.path.join(result_dir, "files",
"%s_%s.csv" % (class_label, embedding_slug)))
info.append({
'filename':
"%s_%s.csv" % (class_label, embedding_slug),
'description':
'%s embedding of cells along %s labels' %
(embedding_name, class_label),
'plot_description':
'%s embedding of cells along %s labels (Classes can be identified '
'with their colors and K-means clusters are marked '
'with different shapes)' % (embedding_name, class_label),
})
metric_results.update({
'kmeans_on_%s_%s_adjusted_mutual_info_score' % (embedding_slug, class_label):
adjusted_mutual_info_score(class_names,
clusters,
average_method="arithmetic"),
'kmeans_on_%s_%s_v_measure_score' % (embedding_slug, class_label):
v_measure_score(class_names, clusters),
'embedding_%s_%s_calinski_harabaz_score' % (embedding_slug, class_label):
calinski_harabaz_score(emb, class_names),
'embedding_%s_%s_silhouette_score' % (embedding_slug, class_label):
silhouette_score(emb, class_names)
})
write_csv(pd.DataFrame(info),
os.path.join(result_dir, "files", "info.csv"))
result_path = os.path.join(result_dir, "result.txt")
with open(result_path, 'w') as file:
file.write("## METRICS:\n")
for metric in sorted(metric_results):
file.write("%s\t%4f\n" % (metric, metric_results[metric]))
file.write("##\n## ADDITIONAL INFO:\n")
log("Evaluation results saved to `%s`" % result_path)
if visualization != "none":
self.visualize_result(result_dir, output_type=visualization)
return metric_results
def evaluate_result(self, processed_count_file_path, result_dir,
visualization, **kwargs):
normalization = kwargs['normalization']
transformation = kwargs['transformation']
# Load hidden state and data
count_matrix_lq, original_columns, column_permutation, count_matrix_hq = self._load_hidden_state(
)
# Load imputed data
imputed_data = read_table_file(processed_count_file_path)
# Restore column names and order
imputed_data = rearrange_and_rename_columns(imputed_data,
original_columns,
column_permutation)
# Replace negative values with zero
imputed_data = imputed_data.clip(lower=0)
# Data transformations
imputed_data = transformations[transformation](
normalizations[normalization](imputed_data))
count_matrix_hq = transformations[transformation](
normalizations[normalization](count_matrix_hq))
# Evaluation
rmse_distances = []
mae_distances = []
euclidean_distances = []
cosine_distances = []
correlation_distances = []
for i in range(count_matrix_hq.shape[1]):
non_zeros = np.logical_and(count_matrix_hq.values[:, i] > 0,
count_matrix_lq.values[:, i] == 0)
hq = count_matrix_hq.values[non_zeros, i]
lq = count_matrix_lq.values[non_zeros, i]
y = imputed_data.values[non_zeros, i]
if np.sum(y) > 0:
y = y * np.sum(hq) / np.sum(y)
rmse_distances.append(float(np.mean(np.square(hq - y)**0.5)))
mae_distances.append(float(np.mean(np.abs(hq - y))))
euclidean_distances.append(
pdist(np.vstack((hq, y)), 'euclidean')[0])
cosine_distances.append(pdist(np.vstack((hq, y)), 'cosine')[0])
correlation_distances.append(
pdist(np.vstack((hq, y)), 'correlation')[0])
metric_results = {
'cell_root_mean_squared_error': np.mean(rmse_distances),
'cell_mean_absolute_error': np.mean(mae_distances),
'cell_mean_euclidean_distance': np.mean(euclidean_distances),
'cell_mean_cosine_distance': np.mean(cosine_distances),
'cell_mean_correlation_distance': np.mean(correlation_distances)
}
# Save results to a file
make_sure_dir_exists(os.path.join(result_dir, "files"))
result_path = os.path.join(result_dir, "result.txt")
with open(result_path, 'w') as file:
file.write("## METRICS:\n")
for metric in sorted(metric_results):
file.write("%s\t%4f\n" %
(metric, float(metric_results[metric])))
file.write("##\n## ADDITIONAL INFO:\n")
file.write(
"# CELL\troot_mean_squared_error\tmean_absolute_error\tmean_euclidean_distance\t"
"mean_cosine_distance\tmean_correlation_distance:\n")
for i in range(count_matrix_hq.shape[1]):
file.write(
"# %s\t%f\t%f\t%f\t%f\t%f\n" %
(count_matrix_hq.columns.values[i], rmse_distances[i],
mae_distances[i], euclidean_distances[i],
cosine_distances[i], correlation_distances[i]))
log("Evaluation results saved to `%s`" % result_path)
if visualization != "none":
self.visualize_result(result_dir, output_type=visualization)
return metric_results
def evaluate_result(self, processed_count_file_path, result_dir,
visualization, **kwargs):
make_sure_dir_exists(os.path.join(result_dir, "files"))
transformation = kwargs['transformation']
# Load hidden state and data
_, original_columns, column_permutation, count_adt, protein_rna_mapping = self._load_hidden_state(
)
# Load imputed data
imputed_rna = read_table_file(processed_count_file_path)
# Restore column names and order
imputed_rna = rearrange_and_rename_columns(imputed_rna,
original_columns,
column_permutation)
# Data transformations
imputed_rna = transformations[transformation](imputed_rna)
count_adt = transformations[transformation](count_adt)
# Use related data
adt = count_adt.loc[[
prot for prot in count_adt.index.values
if (protein_rna_mapping[prot] in imputed_rna.index.values)
]].copy()
adt.index = ["prot_" + p for p in adt.index.values]
rna = imputed_rna.loc[[
protein_rna_mapping[prot] for prot in count_adt.index.values
if (protein_rna_mapping[prot] in imputed_rna.index.values)
]]
rna.index = ["gene_" + g for g in rna.index.values]
info = []
write_csv(adt, os.path.join(result_dir, "files", "adt.csv"))
info.append({
'filename':
"adt.csv",
'description':
'Protein expressions (adt) after transformation',
'plot_description':
'Protein expressions (adt) after transformation',
})
write_csv(rna, os.path.join(result_dir, "files", "rna.csv"))
info.append({
'filename':
"rna.csv",
'description':
'Gene expressions of genes related to adt data after transformation',
'plot_description':
'Gene expressions of genes related to adt data after transformation',
})
n = adt.shape[0]
# Calculating Spearman correlations
combined_df = pd.concat((adt, rna)).transpose()
correlations = combined_df.corr(method="spearman")
adt_adt_spearmanr = correlations.iloc[:n, :n]
rna_rna_spearmanr = correlations.iloc[n:, n:]
adt_rna_spearmanr = correlations.iloc[:n, n:]
write_csv(
correlations,
os.path.join(result_dir, "files", "spearman_correlations.csv"))
info.append({
'filename':
"spearman_correlations.csv",
'description':
'Pairwise Spearman correlations (first n items are '
'adt expressions and second n items are rna expressions)',
'plot_description':
'Pairwise Spearman correlations (first n items are '
'adt expressions and second n items are rna expressions)',
})
# Calculating Pearson correlations
combined_df = pd.concat((adt, rna)).transpose()
correlations = combined_df.corr(method="pearson")
adt_adt_pearsonr = correlations.iloc[:n, :n]
rna_rna_pearsonr = correlations.iloc[n:, n:]
adt_rna_pearsonr = correlations.iloc[:n, n:]
write_csv(
correlations,
os.path.join(result_dir, "files", "pearson_correlations.csv"))
info.append({
'filename':
"pearson_correlations.csv",
'description':
'Pairwise Pearson correlations (first n items are '
'adt expressions and second n items are rna expressions)',
'plot_description':
'Pairwise Pearson correlations (first n items are '
'adt expressions and second n items are rna expressions)',
})
# Evaluation
metric_results = {
'rna_protein_mean_spearman_correlatoin':
np.mean(adt_rna_spearmanr.values.diagonal()),
'rna_protein_mean_pearson_correlatoin':
np.mean(adt_rna_pearsonr.values.diagonal()),
'MSE_of_adt_adt_and_rna_rna_spearman_correlations':
np.mean((adt_adt_spearmanr.values - rna_rna_spearmanr.values)**2),
'MSE_of_adt_adt_and_rna_rna_pearson_correlations':
np.mean((adt_adt_pearsonr.values - rna_rna_pearsonr.values)**2)
}
write_csv(pd.DataFrame(info),
os.path.join(result_dir, "files", "info.csv"))
# Save results to a file
result_path = os.path.join(result_dir, "result.txt")
with open(result_path, 'w') as file:
file.write("## METRICS:\n")
for metric in sorted(metric_results):
file.write("%s\t%4f\n" %
(metric, float(metric_results[metric])))
file.write("##\n## ADDITIONAL INFO:\n")
file.write("## Pearson of adt/rna:\n")
file.write("## " +
"\n## ".join(adt_rna_pearsonr.to_string().split("\n")) +
"\n")
file.write('## Spearman of adt/rna:\n')
file.write(
"## " +
"\n## ".join(adt_rna_spearmanr.to_string().split("\n")) + "\n")
file.write("## Pearson of adt/adt:\n")
file.write("## " +
"\n## ".join(adt_adt_pearsonr.to_string().split("\n")) +
"\n")
file.write("## Pearson of rna/rna:\n")
file.write("## " +
"\n## ".join(rna_rna_pearsonr.to_string().split("\n")) +
"\n")
file.write('## Spearman of adt/adt:\n')
file.write(
"## " +
"\n## ".join(adt_adt_spearmanr.to_string().split("\n")) + "\n")
file.write('## Spearman of rna/rna:\n')
file.write(
"## " +
"\n## ".join(rna_rna_spearmanr.to_string().split("\n")) + "\n")
log("Evaluation results saved to `%s`" % result_path)
if visualization != "none":
self.visualize_result(result_dir, output_type=visualization)
return metric_results
def _download_data_set(self):
make_sure_dir_exists(os.path.dirname(self.DATA_SET_FILE_PATH))
download_file_if_not_exists(self.DATA_SET_URL,
self.DATA_SET_FILE_PATH,
self.DATA_SET_MD5_SUM)
def evaluate_result(self, processed_count_file, result_dir, visualization,
**kwargs):
normalization = kwargs['normalization']
transformation = kwargs['transformation']
clear_cache = kwargs['clear_cache']
make_sure_dir_exists(os.path.join(result_dir, "files"))
info = []
data, imputed_data = self._load_data_and_imputed_data_for_evaluation(
processed_count_file)
gold_standard_classes = [
column_name.split("_")[0] for column_name in data.columns.values
]
# Data transformations
if np.sum(imputed_data.values < 0) > 0:
log("Observed some negative values!")
imputed_data[imputed_data < 0] = 0
imputed_data = transformations[transformation](
normalizations[normalization](imputed_data))
G1_S_related_part_of_imputed_data, G2_M_related_part_of_imputed_data = self._get_related_part(
imputed_data)
related_part_of_imputed_data = pd.concat([
G1_S_related_part_of_imputed_data,
G2_M_related_part_of_imputed_data
])
write_csv(
G1_S_related_part_of_imputed_data,
os.path.join(result_dir, "files",
"G1_S_related_part_of_imputed_data.csv"))
info.append({
'filename': 'G1_S_related_part_of_imputed_data.csv',
'description': 'Vales of genes related to G1/S',
'plot_description': 'Heatmap of Genes related to G1/S',
})
write_csv(
G2_M_related_part_of_imputed_data,
os.path.join(result_dir, "files",
"G2_M_related_part_of_imputed_data.csv"))
info.append({
'filename': 'G2_M_related_part_of_imputed_data.csv',
'description': 'Vales of genes related to G2/M',
'plot_description': 'Heatmap of Genes related to G2/M',
})
svm_results, knn_results = self._get_classification_results(
related_part_of_imputed_data, gold_standard_classes)
embedded_data_file_path = os.path.join(result_dir, "files",
"embedded_data.pkl.gz")
if os.path.exists(embedded_data_file_path) and not clear_cache:
embedded_data = load_gzip_pickle(embedded_data_file_path)
else:
embedded_data = self._get_embeddings(related_part_of_imputed_data)
dump_gzip_pickle(embedded_data, embedded_data_file_path)
metric_results = {
"classification_svm_mean_accuracy": np.mean(svm_results),
"classification_knn_mean_accuracy": np.mean(knn_results)
}
embedded_data["identity"] = related_part_of_imputed_data.transpose()
for i, embedding_name in enumerate(embedded_data):
emb = embedded_data[embedding_name]
k_means = KMeans(n_clusters=3)
k_means.fit(emb)
clusters = k_means.predict(emb)
embedding_slug = embedding_name.replace(" ", "_").lower()
if embedding_name != "identity":
embedding_df = pd.DataFrame(
{
"X": emb[:, 0],
"Y": emb[:, 1],
"class": gold_standard_classes,
"k_means_clusters": clusters
},
index=data.columns.values)
write_csv(
embedding_df,
os.path.join(result_dir, "files",
"%s.csv" % embedding_slug))
info.append({
'filename':
"%s.csv" % embedding_slug,
'description':
'%s embedding of cells considering genes related '
'to cell-cycle' % embedding_name,
'plot_description':
'%s embedding of cells considering genes related '
'to cell-cycle (K-means clusters are marked '
'with different shapes)' % embedding_name,
})
metric_results.update({
'kmeans_on_%s_adjusted_mutual_info_score' % embedding_slug:
adjusted_mutual_info_score(gold_standard_classes,
clusters,
average_method="arithmetic"),
'kmeans_on_%s_v_measure_score' % embedding_slug:
v_measure_score(gold_standard_classes, clusters),
'embedding_%s_calinski_harabaz_score' % embedding_slug:
calinski_harabaz_score(emb, gold_standard_classes),
'embedding_%s_silhouette_score' % embedding_slug:
silhouette_score(emb, gold_standard_classes)
})
write_csv(pd.DataFrame(info),
os.path.join(result_dir, "files", "info.csv"))
result_path = os.path.join(result_dir, "result.txt")
with open(result_path, 'w') as file:
file.write("## METRICS:\n")
for metric in sorted(metric_results):
file.write("%s\t%4f\n" % (metric, metric_results[metric]))
file.write("##\n## ADDITIONAL INFO:\n")
file.write("## SVM classifiers accuracies: %s\n" %
str(svm_results))
file.write("## KNN classifiers accuracies: %s\n" %
str(knn_results))
log("Evaluation results saved to `%s`" % result_path)
if visualization != "none":
self.visualize_result(result_dir, output_type=visualization)
return metric_results
def evaluate_result(self, processed_count_file_path, result_dir,
visualization, **kwargs):
transformation = kwargs['transformation']
make_sure_dir_exists(os.path.join(result_dir, "files"))
# Load hidden state and data
scaled_data, original_columns, column_permutation = self._load_hidden_state(
)
# Load imputed data
imputed_data = read_table_file(processed_count_file_path)
# Restore column names and order
imputed_data = rearrange_and_rename_columns(imputed_data,
original_columns,
column_permutation)
# Replace negative values with zero
imputed_data = imputed_data.clip(lower=0)
# Data transformation
scaled_data = transformations[transformation](scaled_data)
imputed_data = transformations[transformation](imputed_data)
# Evaluation
rmse_distances = []
mae_distances = []
euclidean_distances = []
cosine_distances = []
correlation_distances = []
rmse = float(
np.sum(
np.where(scaled_data.values > 0, 1, 0) *
np.square(scaled_data.values - imputed_data.values)) /
np.sum(np.where(scaled_data.values > 0, 1, 0)))**0.5
mae = float(
np.sum(
np.where(scaled_data.values > 0, 1, 0) *
np.abs(scaled_data.values - imputed_data.values)) /
np.sum(np.where(scaled_data.values > 0, 1, 0)))
for i in range(scaled_data.shape[1]):
non_zeros = scaled_data.values[:, i] > 0
x = scaled_data.values[non_zeros, i]
y = imputed_data.values[non_zeros, i]
rmse_distances.append(
float(np.sum(np.square(x - y)) / np.sum(non_zeros))**0.5)
mae_distances.append(
float(np.sum(np.abs(x - y)) / np.sum(non_zeros)))
cosine_distances.append(pdist(np.vstack((x, y)), 'cosine')[0])
euclidean_distances.append(
pdist(np.vstack((x, y)), 'euclidean')[0])
correlation_distances.append(
pdist(np.vstack((x, y)), 'correlation')[0])
metric_results = {
'all_mean_absolute_error_on_non_zeros':
mae,
'all_root_mean_squared_error_on_non_zeros':
rmse,
'cell_mean_mean_absolute_error_on_non_zeros':
np.mean(mae_distances),
'cell_mean_root_mean_squared_error_on_non_zeros':
np.mean(rmse_distances),
'cell_mean_euclidean_distance':
np.mean(euclidean_distances),
'cell_mean_cosine_distance':
np.mean(cosine_distances),
'cell_mean_correlation_distance':
np.mean(correlation_distances),
}
# Save results to a file
result_path = os.path.join(result_dir, "result.txt")
with open(result_path, 'w') as file:
file.write("## METRICS:\n")
for metric in sorted(metric_results):
file.write("%s\t%4f\n" %
(metric, float(metric_results[metric])))
file.write("##\n## ADDITIONAL INFO:\n")
file.write(
"# CELL\troot_mean_squared_error_on_non_zeros\tmean_absolute_error_on_non_zeros\t"
"euclidean_distance_on_non_zeros\tcosine_distance_on_non_zeros\tcorrelation_distance_on_non_zeros:\n"
)
for i in range(scaled_data.shape[1]):
file.write("# %s\t%f\t%f\t%f\t%f\t%f\n" %
(scaled_data.columns.values[i], rmse_distances[i],
mae_distances[i], euclidean_distances[i],
cosine_distances[i], correlation_distances[i]))
log("Evaluation results saved to `%s`" % result_path)
if visualization != "none":
self.visualize_result(result_dir, output_type=visualization)
return metric_results
def evaluate_result(self, processed_count_file_path, result_dir,
visualization, **kwargs):
make_sure_dir_exists(os.path.join(result_dir, "files"))
info = []
# Load hidden state and data
data, mask, original_columns, column_permutation = self._load_hidden_state(
)
# Load imputed data
imputed_data = read_table_file(processed_count_file_path)
# Restore column names and order
imputed_data = rearrange_and_rename_columns(imputed_data,
original_columns,
column_permutation)
# Replace negative values with zero
imputed_data = imputed_data.clip(lower=0)
# Evaluation
log_diff = np.abs(transformations["log"](data) -
transformations["log"](imputed_data))
sqrt_diff = np.abs(transformations["sqrt"](data) -
transformations["sqrt"](imputed_data))
mse_on_log = float(
np.sum(
np.sum(mask * np.where(data != 0, 1, 0) * np.square(log_diff)))
/ np.sum(np.sum(mask * np.where(data != 0, 1, 0))))
mae_on_log = float(
np.sum(np.sum(mask * np.where(data != 0, 1, 0) * np.abs(log_diff)))
/ np.sum(np.sum(mask * np.where(data != 0, 1, 0))))
mse_on_sqrt = float(
np.sum(
np.sum(
mask * np.where(data != 0, 1, 0) * np.square(sqrt_diff))) /
np.sum(np.sum(mask * np.where(data != 0, 1, 0))))
mae_on_sqrt = float(
np.sum(np.sum(
mask * np.where(data != 0, 1, 0) * np.abs(sqrt_diff))) /
np.sum(np.sum(mask * np.where(data != 0, 1, 0))))
metric_results = {
'RMSE_sqrt': mse_on_sqrt**0.5,
'MAE_sqrt': mae_on_sqrt,
'RMSE_log': mse_on_log**0.5,
'MAE_log': mae_on_log
}
masked_locations = []
mask_values = mask.values
for x in range(mask_values.shape[0]):
for y in range(mask_values.shape[1]):
if mask_values[x, y] == 1:
masked_locations.append((x, y))
original_values = []
predicted_values = []
for (x, y) in masked_locations:
original_values.append(data.iloc[x, y])
predicted_values.append(imputed_data.iloc[x, y])
original_values = np.asarray(original_values)
predicted_values = np.asarray(predicted_values)
predictions_df = pd.DataFrame({
'original': original_values,
'predicted': predicted_values
})
write_csv(predictions_df,
os.path.join(result_dir, "files", "predictions.csv"))
info.append({
'filename':
"predictions.csv",
'description':
'Original masked values along predicted values',
'plot_description':
'Predicted values vs. original masked values',
})
write_csv(pd.DataFrame(info),
os.path.join(result_dir, "files", "info.csv"))
# Save results to a file
result_path = os.path.join(result_dir, "result.txt")
with open(result_path, 'w') as file:
file.write("## METRICS:\n")
for metric in sorted(metric_results):
file.write("%s\t%4f\n" % (metric, metric_results[metric]))
file.write("##\n## ADDITIONAL INFO:\n")
file.write("# GENE\tCELL\tGOLD_STANDARD\tRESULT:\n")
for (x, y) in masked_locations:
file.write("# %s\t%s\t%f\t%f\n" %
(data.index.values[x], data.columns.values[y],
data.iloc[x, y], imputed_data.iloc[x, y]))
log("Evaluation results saved to `%s`" % result_path)
if visualization != "none":
self.visualize_result(result_dir, output_type=visualization)
return metric_results
