def continue_run_rand_nn(matrixA, matrixB, num_gene, num_pathway, layer1,
layer2, layer3, path, dir_opt, RNA_seq_filename,
input_num, epoch, batch_size, verbose):
# RECONSTRCUT TO BE TRAINED MODEL
model = RandNN().keras_rand_nn(matrixA, matrixB, num_gene, layer0, layer1,
layer2, layer3)
with open(path + '/layer_bias_list.txt', 'rb') as filebias:
layer_bias_list = pickle.load(filebias)
with open(path + '/layer_weight_list.txt', 'rb') as fileweight:
layer_weight_list = pickle.load(fileweight)
model.compile(loss='mean_squared_error',
optimizer='adam',
metrics=['mse', 'accuracy'])
xTmp, yTmp = LoadData(dir_opt, RNA_seq_filename).load_train(0, 1)
model.fit(xTmp, yTmp, epochs=1, validation_split=1, verbose=0)
model_layer_list = []
num_layer = len(model.layers)
for i in range(num_layer):
each_layer_list = [layer_weight_list[i], layer_bias_list[i]]
model_layer_list.append(each_layer_list)
model.layers[i].set_weights(each_layer_list)
# AUTO UPDATE WEIGHT
model, history, num_layer, path = RunRandNN(
model, dir_opt, RNA_seq_filename).train(input_num, epoch, batch_size,
verbose)
return model, history, path
def manual_test_rand_nn(matrixA, matrixB, num_gene, num_pathway, layer1,
layer2, layer3, path, dir_opt, RNA_seq_filename):
# RECONSTRCUT TEST MODEL
model = RandNN().keras_rand_nn(matrixA, matrixB, num_gene, layer0, layer1,
layer2, layer3)
with open(path + '/layer_bias_list.txt', 'rb') as filebias:
layer_bias_list = pickle.load(filebias)
with open(path + '/layer_weight_list.txt', 'rb') as fileweight:
layer_weight_list = pickle.load(fileweight)
model.compile(loss='mean_squared_error',
optimizer='adam',
metrics=['mse', 'accuracy'])
xTmp, yTmp = LoadData(dir_opt, RNA_seq_filename).load_train(0, 1)
model.fit(xTmp, yTmp, epochs=1, validation_split=1, verbose=0)
model_layer_list = []
num_layer = len(model.layers)
for i in range(num_layer):
each_layer_list = [layer_weight_list[i], layer_bias_list[i]]
model_layer_list.append(each_layer_list)
model.layers[i].set_weights(each_layer_list)
# PREDICT MODEL USING [xTe, yTe]
verbose = 1
y_pred, score = RunRandNN(model, dir_opt,
RNA_seq_filename).test(verbose, path)
def gene_analysis(self, RNA_seq_filename, pathway_filename, path,
post_data_path, epoch_time):
# ANALYSE PATHWAY GENE AND DISTRIBUTION
dir_opt = self.dir_opt
zero_final_dl_input_df, input_num, num_feature, cellline_gene_df, num_gene, num_pathway = LoadData(
dir_opt, RNA_seq_filename).pre_load_train()
layer0 = num_pathway
layer1 = 256
layer2 = 128
layer3 = 32
matrixA = GenMatrix(dir_opt, RNA_seq_filename,
pathway_filename).feature_gene_matrix(
num_feature, num_gene)
matrixB = GenMatrix(dir_opt, RNA_seq_filename,
pathway_filename).gene_pathway_matrix(num_pathway)
# RECONSTRUCT THE DECOMPOSED MODEL
input_model, gene_model, pathway_model, model = RandNN().keras_rand_nn(
matrixA, matrixB, num_gene, num_pathway, layer1, layer2, layer3)
# AUTO REBUILD THE MODEL
# model = model.load_weights(path + '/model.h5')
# MANUAL REBUILD THE MODEL (BETTER)
with open(path + '/layer_list.txt', 'rb') as filelayer:
layer_list = pickle.load(filelayer)
model.compile(loss='mean_squared_error',
optimizer='adam',
metrics=['mse', 'accuracy'])
xTmp, yTmp = LoadData(dir_opt, RNA_seq_filename).load_train(0, 1)
model.fit(xTmp, yTmp, epochs=1, validation_split=1, verbose=0)
num_layer = len(model.layers)
for i in range(num_layer):
model.get_layer(index=i).set_weights(layer_list[i])
# GENE IMPORTANT ANALYSIS
# ON TRAINING DATA SET
xTr = np.load(post_data_path + '/xTr.npy')
gene_x = np.array(input_model.predict(xTr))
pathway_analyzer = innvestigate.create_analyzer(
'smoothgrad',
gene_model,
noise_scale=(np.max(gene_x) - np.min(gene_x)) * 0.1)
analysis = pathway_analyzer.analyze(gene_x)
mean_analysis = np.mean(analysis, axis=0)
print('Gene Analysis')
print(mean_analysis.shape)
# MAKE BARPOLT ON GENES IMPORTANCE
top = 50
mean_analysis = np.absolute(mean_analysis)
# GET AND SAVE TOP GENE INDEX
top_gene_index = heapq.nlargest(top, range(len(mean_analysis)),
mean_analysis.take)
np.save(path + '/train_top_gene_index' + epoch_time + '.npy',
top_gene_index)
top_gene_value = mean_analysis[top_gene_index]
cellline_df = pd.read_csv('.' + dir_opt + '/filtered_data/' +
RNA_seq_filename + '.csv')
gene_name_list = cellline_df['geneSymbol']
gene_name_dict = {k: v for k, v in enumerate(gene_name_list)}
top_gene_name = [gene_name_dict.get(key) for key in top_gene_index]
top_gene_name_value_dict = dict(zip(top_gene_name, top_gene_value))
# BARPLOT ON GENES AND SAVE
plt.figure(figsize=(16, 9))
plt.bar(range(len(top_gene_name_value_dict)),
list(top_gene_name_value_dict.values()))
plt.xticks(range(len(top_gene_name_value_dict)),
list(top_gene_name_value_dict.keys()),
rotation=30,
fontsize=10,
ha='right')
plt.title(
'Top 50 Genes With Largest Absolute Importance Scores of 1684 Genes On Training Dataset',
fontsize=16)
plt.tight_layout()
# SAVE TRAINING PLOT FIGURE
file_name = 'epoch_' + epoch_time + '_train_gene_barplot'
train_path = '.' + dir_opt + '/plot/%s' % (file_name) + '.png'
unit = 1
while os.path.exists(train_path):
train_path = '.' + dir_opt + '/plot/%s_%d' % (file_name,
unit) + '.png'
unit += 1
plt.savefig(train_path, dpi=300)
# ON TEST DATA SET
xTe = np.load(post_data_path + '/xTe.npy')
gene_x = np.array(input_model.predict(xTe))
gene_analyzer = innvestigate.create_analyzer(
'smoothgrad',
gene_model,
noise_scale=(np.max(gene_x) - np.min(gene_x)) * 0.1)
analysis = gene_analyzer.analyze(gene_x)
mean_analysis = np.mean(analysis, axis=0)
print('Gene Analysis')
print(mean_analysis.shape)
# MAKE BARPOLT ON GENES IMPORTANCE
top = 50
mean_analysis = np.absolute(mean_analysis)
# GET AND SAVE TOP GENE INDEX
top_gene_index = heapq.nlargest(top, range(len(mean_analysis)),
mean_analysis.take)
np.save(path + '/test_top_gene_index' + epoch_time + '.npy',
top_gene_index)
top_gene_value = mean_analysis[top_gene_index]
cellline_df = pd.read_csv('.' + dir_opt + '/filtered_data/' +
RNA_seq_filename + '.csv')
gene_name_list = cellline_df['geneSymbol']
gene_name_dict = {k: v for k, v in enumerate(gene_name_list)}
top_gene_name = [gene_name_dict.get(key) for key in top_gene_index]
top_gene_name_value_dict = dict(zip(top_gene_name, top_gene_value))
# BARPLOT ON GENES AND SAVE
plt.figure(figsize=(16, 9))
plt.bar(range(len(top_gene_name_value_dict)),
list(top_gene_name_value_dict.values()))
plt.xticks(range(len(top_gene_name_value_dict)),
list(top_gene_name_value_dict.keys()),
rotation=30,
fontsize=10,
ha='right')
plt.title(
'Top 50 Genes With Largest Absolute Importance Scores of 1684 Genes On Test Dataset',
fontsize=16)
plt.tight_layout()
# SAVE TEST PLOT FIGURE
file_name = 'epoch_' + epoch_time + '_test_gene_barplot'
test_path = '.' + dir_opt + '/plot/%s' % (file_name) + '.png'
unit = 1
while os.path.exists(test_path):
test_path = '.' + dir_opt + '/plot/%s_%d' % (file_name,
unit) + '.png'
unit += 1
plt.savefig(test_path, dpi=300)
def pathway_analysis(self, RNA_seq_filename, pathway_filename, path,
post_data_path, epoch_time):
# ANALYSE PATHWAY GENE AND DISTRIBUTION
dir_opt = self.dir_opt
train_input_df, input_num, num_feature, rna_df, cpnum_df, num_gene, num_pathway = LoadData(
dir_opt).pre_load_train()
layer0 = num_pathway
layer1 = 256
layer2 = 128
layer3 = 32
matrixA = GenMatrix.feature_gene_matrix(num_feature, num_gene)
matrixB = GenMatrix.gene_pathway_matrix()
# RECONSTRUCT THE DECOMPOSED MODEL
input_model, gene_model, pathway_model, model = RandNN().keras_rand_nn(
matrixA, matrixB, num_gene, num_pathway, layer1, layer2, layer3)
# AUTO REBUILD THE MODEL
# model = model.load_weights(path + '/model.h5')
# MANUAL REBUILD THE MODEL (BETTER)
with open(path + '/layer_list.txt', 'rb') as filelayer:
layer_list = pickle.load(filelayer)
model.compile(loss='mean_squared_error',
optimizer='adam',
metrics=['mse', 'accuracy'])
xTmp, yTmp = LoadData(dir_opt).load_train(0, 1)
model.fit(xTmp, yTmp, epochs=1, validation_split=1, verbose=0)
num_layer = len(model.layers)
for i in range(num_layer):
model.get_layer(index=i).set_weights(layer_list[i])
# RE-LOAD EACH MODEL
input_model = model.layers[1]
gene_model = model.layers[2]
pathway_model = model.layers[3]
# GENE IMPORTANT ANALYSIS
# ON TEST DATA ONLY
xTe = np.load(post_data_path + '/xTe.npy')
gene_x = np.array(input_model.predict(xTe))
# PATHWAY INPORTANCE ANALYSIS ON TEST INDEX
# import pdb; pdb.set_trace()
pathway_x = gene_model.predict(gene_x)
gene_pathway_df = pd.read_table('.' + dir_opt + '/init_data/' +
pathway_filename + '.txt')
pathway_name_list = list(gene_pathway_df.columns[1:])
pathway_analyzer = innvestigate.create_analyzer(
"smoothgrad",
pathway_model,
noise_scale=(np.max(pathway_x) - np.min(pathway_x)) * 0.1)
analysis = pathway_analyzer.analyze(pathway_x)
print('PATHWAY IMPORTANCE ANALYSIS...')
file_name = 'epoch_' + epoch_time + '_pathway_analysis'
analysis_path = '.' + dir_opt + '/plot/%s' % (file_name) + '.pdf'
unit = 1
while os.path.exists(analysis_path):
analysis_path = '.' + dir_opt + '/plot/%s_%d' % (file_name,
unit) + '.pdf'
unit += 1
pdf = PdfPages(analysis_path)
plt.figure(figsize=(8, 6))
ax = plt.imshow(analysis.squeeze(),
cmap='seismic',
interpolation='nearest',
aspect="auto")
cb = plt.colorbar(ax)
cb.ax.tick_params(labelsize=8)
plt.ylabel('Sample Index')
plt.xlabel('Pathway')
plt.xticks(rotation=45)
plt.tick_params(labelsize=8)
pdf.savefig()
plt.close()
pdf.close()
# PATHWAY DISTRIBUTION ANALYSIS ON TEST INDEX
print('PATHWAY DISTRIBUTION ANALYSIS...')
file_name = 'epoch_' + epoch_time + '_pathway_distribution'
distribution_path = '.' + dir_opt + '/plot/%s' % (file_name) + '.pdf'
unit = 1
while os.path.exists(distribution_path):
distribution_path = '.' + dir_opt + '/plot/%s_%d' % (file_name,
unit) + '.pdf'
unit += 1
pdf = PdfPages(distribution_path)
pathway_name = pathway_name_list
fig = plt.figure(figsize=(15, 20))
for i in range(num_pathway):
ax = fig.add_subplot(11, 5, i + 1)
plt.xlim([-0.1, 0.1])
sns.kdeplot(analysis[:, i], shade=True)
plt.title(pathway_name[i], fontsize='small', fontweight='bold')
plt.tight_layout()
plt.subplots_adjust(wspace=0.5, hspace=0.5)
pdf.savefig()
plt.close()
pdf.close()
def build_rand_nn(matrixA, matrixB, num_gene, num_pathway, layer1, layer2,
layer3):
input_model, gene_model, pathway_model, model = RandNN().keras_rand_nn(
matrixA, matrixB, num_gene, num_pathway, layer1, layer2, layer3)
return input_model, gene_model, pathway_model, model
def build_rand_nn(matrixA, matrixB, num_gene, layer0, layer1, layer2, layer3):
model = RandNN().keras_rand_nn(matrixA, matrixB, num_gene, layer0, layer1,
layer2, layer3)
return model