def train_test_rf_vary_data_size(prefix,
motif_scoring_kwargs=None,
X_train=None,
y_train=None,
X_valid=None,
y_valid=None,
X_test=None,
y_test=None,
train_set_sizes=None):
motif_scores_train = get_motif_scores(X_train, **motif_scoring_kwargs)
motif_scores_test = get_motif_scores(X_test, **motif_scoring_kwargs)
rf_results = []
for train_set_size in train_set_sizes:
ofname_infix = dict2string(motif_scoring_kwargs)
ofname_infix = "%s.train_set_size_%s" % (ofname_infix,
str(train_set_size))
ofname = "%s.%s.rf.pkl" % (prefix, ofname_infix)
try:
with open(ofname, 'rb') as fp:
rf = pickle.load(fp)
except:
logger.info("training with %i examples.." % (train_set_size))
rf = RandomForest()
rf.train(motif_scores_train[:train_set_size],
y_train[:train_set_size].squeeze())
with open(ofname, 'wb') as fid:
pickle.dump(rf, fid)
rf_results.append(rf.test(motif_scores_test, y_test))
return rf_results
def plot_SequenceDNN_layer_outputs(dnn, simulation_data):
# define layer out functions
import theano
get_conv_output = theano.function([dnn.model.layers[0].input],
dnn.model.layers[0].get_output(train=False),
allow_input_downcast=True)
get_conv_relu_output = theano.function([dnn.model.layers[0].input],
dnn.model.layers[1].get_output(train=False),
allow_input_downcast=True)
get_maxpool_output = theano.function([dnn.model.layers[0].input],
dnn.model.layers[-4].get_output(train=False),
allow_input_downcast=True)
# get layer outputs for a positive simulation example
pos_indx = np.where(simulation_data.y_valid==1)[0][0]
pos_X = simulation_data.X_valid[pos_indx:(pos_indx+1)]
conv_outputs = get_conv_output(pos_X).squeeze()
conv_relu_outputs = get_conv_relu_output(pos_X).squeeze()
maxpool_outputs = get_maxpool_output(pos_X).squeeze()
# plot layer outputs
fig = plt.figure(figsize=(15, 12))
ax1 = fig.add_subplot(3, 1, 3)
heatmap = ax1.imshow(conv_outputs, aspect='auto', interpolation='None', cmap='seismic')
fig.colorbar(heatmap)
ax1.set_ylabel("Convolutional Filters")
ax1.set_xlabel("Position")
ax1.get_yaxis().set_ticks([])
ax1.get_xaxis().set_ticks([])
ax1.set_title("SequenceDNN outputs from convolutional layer.\t\
Locations of motif sites are highlighted in grey.")
ax2 = fig.add_subplot(3, 1, 2)
heatmap = ax2.imshow(conv_relu_outputs, aspect='auto', interpolation='None', cmap='seismic')
fig.colorbar(heatmap)
ax2.set_ylabel("Convolutional Filters")
ax2.get_yaxis().set_ticks([])
ax2.get_xaxis().set_ticks([])
ax2.set_title("Convolutional outputs after ReLU transformation.\t\
Locations of motif sites are highlighted in grey.")
ax3 = fig.add_subplot(3, 1, 1)
heatmap = ax3.imshow(maxpool_outputs, aspect='auto', interpolation='None', cmap='seismic')
fig.colorbar(heatmap)
ax3.set_title("DNN outputs after max pooling")
ax3.set_ylabel("Convolutional Filters")
ax3.get_yaxis().set_ticks([])
ax3.get_xaxis().set_ticks([])
# highlight motif sites
motif_scores = get_motif_scores(pos_X, simulation_data.motif_names)
motif_sites = [np.argmax(motif_scores[0, i, :]) for i in [0, 1]]
for motif_site in motif_sites:
conv_output_start = motif_site - max(dnn.conv_width-10, 0)
conv_output_stop = motif_site + max(dnn.conv_width-10, 0)
ax1.axvspan(conv_output_start, conv_output_stop, color='grey', alpha=0.5)
ax2.axvspan(conv_output_start, conv_output_stop, color='grey', alpha=0.5)
def train_test_rf_vary_data_size(prefix, motif_scoring_kwargs=None,
X_train=None, y_train=None,
X_valid=None, y_valid=None,
X_test=None, y_test=None,
train_set_sizes=None):
motif_scores_train = get_motif_scores(X_train, **motif_scoring_kwargs)
motif_scores_test = get_motif_scores(X_test, **motif_scoring_kwargs)
rf_results = []
for train_set_size in train_set_sizes:
ofname_infix = dict2string(motif_scoring_kwargs)
ofname_infix = "%s.train_set_size_%s" % (ofname_infix, str(train_set_size))
ofname = "%s.%s.rf.pkl" % (prefix, ofname_infix)
try:
with open(ofname, 'rb') as fp:
rf = pickle.load(fp)
except:
logger.info("training with %i examples.." % (train_set_size))
rf = RandomForest()
rf.train(motif_scores_train[:train_set_size], y_train[:train_set_size].squeeze())
with open(ofname, 'wb') as fid:
pickle.dump(rf, fid)
rf_results.append(rf.test(motif_scores_test, y_test))
return rf_results
def interpret_data_with_SequenceDNN(dnn, simulation_data):
# get a positive and a negative example from the simulation data
pos_indx = np.where(simulation_data.y_valid == 1)[0][2]
neg_indx = np.where(simulation_data.y_valid == 0)[0][2]
pos_X = simulation_data.X_valid[pos_indx:(pos_indx + 1)]
neg_X = simulation_data.X_valid[neg_indx:(neg_indx + 1)]
# get motif scores, ISM scores, and DeepLIFT scores
scores_dict = defaultdict(OrderedDict)
scores_dict['Positive']['Motif Scores'] = get_motif_scores(
pos_X, simulation_data.motif_names)
scores_dict['Positive']['ISM Scores'] = dnn.in_silico_mutagenesis(
pos_X).max(axis=-2)
scores_dict['Positive']['DeepLIFT Scores'] = dnn.deeplift(pos_X).max(
axis=-2)
scores_dict['Negative']['Motif Scores'] = get_motif_scores(
neg_X, simulation_data.motif_names)
scores_dict['Negative']['ISM Scores'] = dnn.in_silico_mutagenesis(
neg_X).max(axis=-2)
scores_dict['Negative']['DeepLIFT Scores'] = dnn.deeplift(neg_X).max(
axis=-2)
# get motif site locations
motif_sites = {}
motif_sites['Positive'] = [
np.argmax(scores_dict['Positive']['Motif Scores'][0, i, :])
for i in range(len(simulation_data.motif_names))
]
motif_sites['Negative'] = [
np.argmax(scores_dict['Negative']['Motif Scores'][0, i, :])
for i in range(len(simulation_data.motif_names))
]
# organize legends
motif_label_dict = {}
motif_label_dict['Motif Scores'] = simulation_data.motif_names
if len(simulation_data.motif_names) == dnn.num_tasks:
motif_label_dict['ISM Scores'] = simulation_data.motif_names
else:
motif_label_dict['ISM Scores'] = [
'_'.join(simulation_data.motif_names)
]
motif_label_dict['DeepLIFT Scores'] = motif_label_dict['ISM Scores']
# plot scores and highlight motif site locations
seq_length = pos_X.shape[-1]
plots_per_row = 2
plots_per_column = 3
ylim_dict = {
'Motif Scores': (-80, 30),
'ISM Scores': (-1.5, 3.0),
'DeepLIFT Scores': (-1.5, 3.0)
}
motif_colors = ['b', 'r', 'c', 'm', 'g', 'k', 'y']
font_size = 12
num_x_ticks = 5
highlight_width = 5
motif_labels_cache = []
f = plt.figure(figsize=(10, 12))
f.subplots_adjust(hspace=0.15, wspace=0.15)
f.set_tight_layout(True)
for j, key in enumerate(['Positive', 'Negative']):
for i, (score_type, scores) in enumerate(scores_dict[key].iteritems()):
ax = f.add_subplot(plots_per_column, plots_per_row,
plots_per_row * i + j + 1)
ax.set_ylim(ylim_dict[score_type])
ax.set_xlim((0, seq_length))
ax.set_frame_on(False)
if j == 0: # put y axis and ticks only on left side
xmin, xmax = ax.get_xaxis().get_view_interval()
ymin, ymax = ax.get_yaxis().get_view_interval()
ax.add_artist(
Line2D((xmin, xmin), (ymin, ymax),
color='black',
linewidth=2))
ax.get_yaxis().tick_left()
for tick in ax.yaxis.get_major_ticks():
tick.label.set_fontsize(font_size / 1.5)
ax.set_ylabel(score_type)
if j > 0: # remove y axes
ax.get_yaxis().set_visible(False)
if i < (plots_per_column - 1): # remove x axes
ax.get_xaxis().set_visible(False)
if i == (plots_per_column - 1): # set x axis and ticks on bottom
ax.set_xticks(seq_length / num_x_ticks *
(np.arange(num_x_ticks + 1)))
xmin, xmax = ax.get_xaxis().get_view_interval()
ymin, ymax = ax.get_yaxis().get_view_interval()
ax.add_artist(
Line2D((xmin, xmax), (ymin, ymin),
color='black',
linewidth=2))
ax.get_xaxis().tick_bottom()
for tick in ax.xaxis.get_major_ticks():
tick.label.set_fontsize(font_size / 1.5)
ax.set_xlabel("Position")
if j > 0 and i < (plots_per_column - 1): # remove all axes
ax.axis('off')
add_legend = False
for _i, motif_label in enumerate(motif_label_dict[score_type]):
if score_type == 'Motif Scores':
scores_to_plot = scores[0, _i, :]
else:
scores_to_plot = scores[0, 0, 0, :]
if motif_label not in motif_labels_cache:
motif_labels_cache.append(motif_label)
add_legend = True
motif_color = motif_colors[motif_labels_cache.index(
motif_label)]
ax.plot(scores_to_plot, label=motif_label, c=motif_color)
if add_legend:
leg = ax.legend(loc=[0, 0.85],
frameon=False,
fontsize=font_size,
ncol=3,
handlelength=-0.5)
for legobj in leg.legendHandles:
legobj.set_color('w')
for _j, text in enumerate(leg.get_texts()):
text_color = motif_colors[motif_labels_cache.index(
motif_label_dict[score_type][_j])]
text.set_color(text_color)
for motif_site in motif_sites[key]:
ax.axvspan(motif_site - highlight_width,
motif_site + highlight_width,
color='grey',
alpha=0.1)
def plot_SequenceDNN_layer_outputs(dnn, simulation_data):
# define layer out functions
get_conv_output = theano.function(
[dnn.model.layers[0].input],
dnn.model.layers[0].get_output(train=False),
allow_input_downcast=True)
get_conv_relu_output = theano.function(
[dnn.model.layers[0].input],
dnn.model.layers[1].get_output(train=False),
allow_input_downcast=True)
get_maxpool_output = theano.function(
[dnn.model.layers[0].input],
dnn.model.layers[-4].get_output(train=False),
allow_input_downcast=True)
# get layer outputs for a positive simulation example
pos_indx = np.where(simulation_data.y_valid == 1)[0][0]
pos_X = simulation_data.X_valid[pos_indx:(pos_indx + 1)]
conv_outputs = get_conv_output(pos_X).squeeze()
conv_relu_outputs = get_conv_relu_output(pos_X).squeeze()
maxpool_outputs = get_maxpool_output(pos_X).squeeze()
# plot layer outputs
fig = plt.figure(figsize=(15, 12))
ax1 = fig.add_subplot(3, 1, 3)
heatmap = ax1.imshow(conv_outputs,
aspect='auto',
interpolation='None',
cmap='seismic')
fig.colorbar(heatmap)
ax1.set_ylabel("Convolutional Filters")
ax1.set_xlabel("Position")
ax1.get_yaxis().set_ticks([])
ax1.get_xaxis().set_ticks([])
ax1.set_title("SequenceDNN outputs from convolutional layer.\t\
Locations of motif sites are highlighted in grey.")
ax2 = fig.add_subplot(3, 1, 2)
heatmap = ax2.imshow(conv_relu_outputs,
aspect='auto',
interpolation='None',
cmap='seismic')
fig.colorbar(heatmap)
ax2.set_ylabel("Convolutional Filters")
ax2.get_yaxis().set_ticks([])
ax2.get_xaxis().set_ticks([])
ax2.set_title("Convolutional outputs after ReLU transformation.\t\
Locations of motif sites are highlighted in grey.")
ax3 = fig.add_subplot(3, 1, 1)
heatmap = ax3.imshow(maxpool_outputs,
aspect='auto',
interpolation='None',
cmap='seismic')
fig.colorbar(heatmap)
ax3.set_title("DNN outputs after max pooling")
ax3.set_ylabel("Convolutional Filters")
ax3.get_yaxis().set_ticks([])
ax3.get_xaxis().set_ticks([])
# highlight motif sites
motif_scores = get_motif_scores(pos_X, simulation_data.motif_names)
motif_sites = [np.argmax(motif_scores[0, i, :]) for i in [0, 1]]
for motif_site in motif_sites:
conv_output_start = motif_site - max(dnn.conv_width - 10, 0)
conv_output_stop = motif_site + max(dnn.conv_width - 10, 0)
ax1.axvspan(conv_output_start,
conv_output_stop,
color='grey',
alpha=0.5)
ax2.axvspan(conv_output_start,
conv_output_stop,
color='grey',
alpha=0.5)
num_epochs = 100
use_deep_CNN = False
use_RNN = False
print('Generating sequences...')
sequences, labels = simulate_single_motif_detection(
'SPI1_disc1', seq_length, num_positives, num_negatives, GC_fraction)
print('One-hot encoding sequences...')
encoded_sequences = one_hot_encode(sequences)
print('Getting motif scores...')
motif_scores = get_motif_scores(encoded_sequences, motif_names=['SPI1_disc1'])
print('Partitioning data into training, validation and test sets...')
X_train, X_test, y_train, y_test = train_test_split(encoded_sequences, labels, test_size=test_fraction)
X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size=validation_fraction)
print('Adding reverse complements...')
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
print('Randomly splitting data into training and test sets...')
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)
def multi_method_interpret(model,
X,
task_idx,
deeplift_score_func,
motif_names=None,
batch_size=200,
target_layer_idx=-2,
num_refs_per_seq=10,
reference="shuffled_ref",
one_hot_func=None,
pfm=None,
GC_fraction=0.4,
generate_plots=True):
"""
Arguments:
model -- keras model object
X -- numpy array with shape (1, 1, n_bases_in_sample,4) or list of FASTA sequences
task_idx -- numerical index (starting with 0) of task to interpet. For a single-tasked model, you should set this to 0
deeplift_score_fun -- scoring function to use with DeepLIFT algorithm.
motif_names -- a list of motif name strings to scan for in the input sequence; if this is unknown, keep the default value of None
batch_size -- number of samples to interpret at once
target_layer_idx -- should be -2 for classification; -1 for regression
reference -- one of 'shuffled_ref','gc_ref','zero_ref'
num_refs_per_seq -- integer indicating number of references to use for each input sequence \
if the reference is set to 'shuffled_ref';if 'zero_ref' or 'gc_ref' is \
used, this argument is ignored.
one_hot_func -- one hot function to use for encoding FASTA string inputs; if the inputs \
are already one-hot-encoded, use the default of None
generate_plots -- default True. Flag to indicate whether or not interpretation plots \
should be generated
Returns:
dictionary with keys 'motif_scan','ism','gradxinput','deeplift'
"""
outputs = dict()
#1) motif scan (if motif_names !=None)
if motif_names is not None:
print("getting 'motif_scan' value")
outputs['motif_scan'] = get_motif_scores(X,
motif_names,
pfm=pfm,
GC_fraction=GC_fraction,
return_positions=True)
else:
outputs['motif_scan'] = None
#2) ISM
print("getting 'ism' value")
outputs['ism'] = in_silico_mutagenesis(model,
X,
task_idx,
target_layer_idx=target_layer_idx)
#3) Input_Grad
print("getting 'input_grad' value")
outputs['input_grad'] = input_grad(model,
X,
target_layer_idx=target_layer_idx)
#4) DeepLIFT
print("getting 'deeplift' value")
outputs['deeplift'] = deeplift(deeplift_score_func,
X,
batch_size=batch_size,
task_idx=task_idx,
num_refs_per_seq=num_refs_per_seq,
reference=reference,
one_hot_func=one_hot_func)
#generate plots
if generate_plots == True:
plot_all_interpretations([outputs], X)
return outputs
def interpret_SequenceDNN_integrative(dnn, simulation_data):
# get a positive and a negative example from the simulation data
pos_indx = np.where(simulation_data.y_valid==1)[0][0]
pos_X = simulation_data.X_valid[pos_indx:(pos_indx+1)]
neg_indx = np.where(simulation_data.y_valid==0)[0][0]
neg_X = simulation_data.X_valid[neg_indx:(neg_indx+1)]
# get motif scores, ISM scores, and DeepLIFT scores
scores_dict = defaultdict(OrderedDict)
scores_dict['Positive']['Motif Scores'] = get_motif_scores(pos_X, simulation_data.motif_names)
scores_dict['Positive']['ISM Scores'] = dnn.in_silico_mutagenesis(pos_X).max(axis=-2)
scores_dict['Positive']['DeepLIFT Scores'] = dnn.deeplift(pos_X).max(axis=-2)
scores_dict['Negative']['Motif Scores'] = get_motif_scores(neg_X, simulation_data.motif_names)
scores_dict['Negative']['ISM Scores'] = dnn.in_silico_mutagenesis(neg_X).max(axis=-2)
scores_dict['Negative']['DeepLIFT Scores'] = dnn.deeplift(neg_X).max(axis=-2)
# get motif site locations
motif_sites = {}
motif_sites['Positive'] = [np.argmax(scores_dict['Positive']['Motif Scores'][0, i, :])
for i in range(len(simulation_data.motif_names))]
motif_sites['Negative'] = [np.argmax(scores_dict['Negative']['Motif Scores'][0, i, :])
for i in range(len(simulation_data.motif_names))]
# organize legends
motif_label_dict = {}
motif_label_dict['Motif Scores'] = simulation_data.motif_names
if len(simulation_data.motif_names) == dnn.num_tasks:
motif_label_dict['ISM Scores'] = simulation_data.motif_names
else:
motif_label_dict['ISM Scores'] = ['_'.join(simulation_data.motif_names)]
motif_label_dict['DeepLIFT Scores'] = motif_label_dict['ISM Scores']
# plot scores and highlight motif site locations
seq_length = dnn.seq_length
plots_per_row = 2
plots_per_column = 3
ylim_dict = {'Motif Scores': (-80, 30), 'ISM Scores': (-1.5, 3.0), 'DeepLIFT Scores': (-1.5, 3.0)}
motif_colors = ['b', 'r', 'c', 'm', 'g', 'k', 'y']
font_size = 12
num_x_ticks = 5
highlight_width = 5
motif_labels_cache = []
f = plt.figure(figsize=(10,12))
f.subplots_adjust(hspace=0.15, wspace=0.15)
f.set_tight_layout(True)
for j, key in enumerate(['Positive', 'Negative']):
for i, (score_type, scores) in enumerate(scores_dict[key].iteritems()):
ax = f.add_subplot(plots_per_column, plots_per_row, plots_per_row*i+j+1)
ax.set_ylim(ylim_dict[score_type])
ax.set_xlim((0, seq_length))
ax.set_frame_on(False)
if j == 0: # put y axis and ticks only on left side
xmin, xmax = ax.get_xaxis().get_view_interval()
ymin, ymax = ax.get_yaxis().get_view_interval()
ax.add_artist(Line2D((xmin, xmin), (ymin, ymax), color='black', linewidth=2))
ax.get_yaxis().tick_left()
for tick in ax.yaxis.get_major_ticks():
tick.label.set_fontsize(font_size/1.5)
ax.set_ylabel(score_type)
if j > 0: # remove y axes
ax.get_yaxis().set_visible(False)
if i < (plots_per_column-1): # remove x axes
ax.get_xaxis().set_visible(False)
if i == (plots_per_column-1): # set x axis and ticks on bottom
ax.set_xticks(seq_length/num_x_ticks*(np.arange(num_x_ticks+1)))
xmin, xmax = ax.get_xaxis().get_view_interval()
ymin, ymax = ax.get_yaxis().get_view_interval()
ax.add_artist(Line2D((xmin, xmax), (ymin, ymin), color='black', linewidth=2))
ax.get_xaxis().tick_bottom()
for tick in ax.xaxis.get_major_ticks():
tick.label.set_fontsize(font_size/1.5)
ax.set_xlabel("Position")
if j>0 and i<(plots_per_column-1): # remove all axes
ax.axis('off')
add_legend = False
for _i, motif_label in enumerate(motif_label_dict[score_type]):
if score_type=='Motif Scores':
scores_to_plot = scores[0, _i, :]
else:
scores_to_plot = scores.squeeze(axis=2)
if motif_label not in motif_labels_cache:
motif_labels_cache.append(motif_label)
add_legend = True
motif_color = motif_colors[motif_labels_cache.index(motif_label)]
ax.plot(scores_to_plot, label=motif_label, c=motif_color)
if add_legend:
leg = ax.legend(loc=[0,0.85], frameon=False, fontsize=font_size,
ncol=3, handlelength=-0.5)
for legobj in leg.legendHandles:
legobj.set_color('w')
for _i, text in enumerate(leg.get_texts()):
text.set_color(motif_color)
for motif_site in motif_sites[key]:
ax.axvspan(motif_site - highlight_width, motif_site + highlight_width,
color='grey', alpha=0.1)
num_epochs = 100
use_deep_CNN = False
use_RNN = False
print('Generating sequences...')
sequences, labels = simulate_single_motif_detection(
'SPI1_disc1', seq_length, num_positives, num_negatives, GC_fraction)
print('One-hot encoding sequences...')
encoded_sequences = one_hot_encode(sequences)
print('Getting motif scores...')
motif_scores = get_motif_scores(encoded_sequences, motif_names=['SPI1_disc1'])
print('Partitioning data into training, validation and test sets...')
X_train, X_test, y_train, y_test = train_test_split(encoded_sequences, labels, test_size=test_fraction)
X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size=validation_fraction)
print('Adding reverse complements...')
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
print('Randomly splitting data into training and test sets...')
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)