def test_generator_multiple_sim():
eligible_cells = ['K562', 'HepG2', 'H1', 'A549', 'HeLa-S3']
eligible_targets = ['DNase', 'CTCF', 'RAD21']
dataset = EpitomeDataset(targets=eligible_targets, cells=eligible_cells)
label_cell_types = ['K562']
eligible_cells.remove(label_cell_types[0])
similarity_matrix = np.ones([2, dataset.get_data(Dataset.TRAIN).shape[1]])
radii = [1, 10]
results = load_data(dataset.get_data(Dataset.TRAIN), ['K562'],
eligible_cells,
dataset.matrix,
dataset.targetmap,
dataset.cellmap,
radii=radii,
mode=Dataset.RUNTIME,
similarity_matrix=similarity_matrix,
similarity_targets=['DNase', 'CTCF'],
indices=np.arange(0, 10))()
li_results = list(results)
# length should include eligible targets and 2* radius for pos and agreement
# for each of the 2 similarity targets
assert (len(li_results[0][0]) == len(eligible_cells) *
(len(eligible_targets) + len(radii) * 4))
def test_train_early_stop_model(self):
tf.random.set_seed(5)
eligible_cells = ['K562', 'HepG2', 'H1']
eligible_targets = ['DNase', 'CTCF']
dataset = EpitomeDataset(targets=eligible_targets,
cells=eligible_cells)
# set all data to ones so it converges quickly
dataset_shape = dataset.get_data(Dataset.ALL).shape
dataset._data = np.ones(dataset_shape)
# create model and train
model = EpitomeModel(dataset, radii=[], max_valid_batches=10)
results1 = model.test(10)
# results should be about random
m = np.mean(results1['preds'])
assert m > 0.4 and m < 0.6
n_steps = 300
_, num_steps, _ = model.train(n_steps, min_delta=sys.maxsize)
assert num_steps < n_steps
results2 = model.test(self.validation_size)
m = np.mean(results2['preds'])
assert m > 0.6 # TODO should be higher
def test_generator_dnase_array():
# should not fail if similarity_targets are just for DNase and is a single array.
# https://github.com/YosefLab/epitome/issues/4
eligible_cells = ['K562', 'HepG2', 'H1', 'A549', 'HeLa-S3']
eligible_targets = ['DNase', 'CTCF', 'RAD21', 'LARP7']
dataset = EpitomeDataset(targets=eligible_targets, cells=eligible_cells)
test_celltypes = ['K562']
eligible_cells.remove(test_celltypes[0])
radii = [1, 10]
# fake data for DNase
similarity_matrix = np.ones(dataset.get_data(Dataset.TRAIN).shape[1])
results = load_data(dataset.get_data(Dataset.TRAIN),
test_celltypes,
eligible_cells,
dataset.matrix,
dataset.targetmap,
dataset.cellmap,
radii,
mode=Dataset.RUNTIME,
similarity_matrix=similarity_matrix,
similarity_targets='DNase',
return_feature_names=True,
indices=np.arange(0, 10))()
li_results = list(results)
# if we reach here, an error was not thrown :)
assert (len(li_results) == 10)
assert (len(li_results[0][0][0]) == 28)
feature_names = li_results[0][1][0]
assert (len(list(filter(lambda x: '_agree' in x,
feature_names))) == len(radii) *
len(eligible_cells))
def test_continous_generator():
eligible_cells = ['K562', 'HepG2', 'H1', 'A549', 'HeLa-S3']
eligible_targets = ['DNase', 'CTCF', 'RAD21', 'LARP7']
dataset = EpitomeDataset(targets=eligible_targets, cells=eligible_cells)
test_celltypes = ['K562']
eligible_cells.remove(test_celltypes[0])
radii = [1, 10]
# fake data for DNase
similarity_matrix = np.ones(dataset.get_data(Dataset.TRAIN).shape[1])
results = load_data(dataset.get_data(Dataset.TRAIN),
test_celltypes,
eligible_cells,
dataset.matrix,
dataset.targetmap,
dataset.cellmap,
radii,
mode=Dataset.RUNTIME,
continuous=True,
similarity_matrix=similarity_matrix,
similarity_targets='DNase',
return_feature_names=True,
indices=np.arange(0, 10))()
li_results = list(results)
feature_names = li_results[0][1][0]
# make sure we don't have the agreement features
assert (len(li_results[0][0][0]) == 20)
assert (len(list(filter(lambda x: '_agree' in x, feature_names))) == 0)
def test_save(self):
out_path = self.tmpFile()
print(out_path)
# generate 100 records for each chromosome
PER_CHR_RECORDS=10
chroms = []
newStarts = []
binSize = 200
for i in range(1,22): # for all 22 chrs
chrom = 'chr' + str(i)
for j in range(1,PER_CHR_RECORDS+1):
chroms.append(chrom)
newStarts.append(j * binSize)
regions_df = pd.DataFrame({'Chromosome': chroms, 'Start': newStarts})
targets = ['DNase', 'CTCF', 'Rad21', 'SCM3', 'IRF3']
cellTypes = ['K562'] * len(targets) + ['A549'] * len(targets) + ['GM12878'] * len(targets)
targets = targets * 3
row_df = pd.DataFrame({'cellType': cellTypes, 'target': targets})
# generate random data of 0/1s
np.random.seed(0)
rand_data = np.random.randint(2, size=(len(row_df), len(regions_df))).astype('i1')
source = 'rand_source'
assembly = 'hg19'
EpitomeDataset.save(out_path,
rand_data,
row_df,
regions_df,
binSize,
assembly,
source,
valid_chrs = ['chr7'],
test_chrs = ['chr8','chr9'])
# load back in dataset and make sure it checks out
dataset = EpitomeDataset(data_dir = out_path)
assert dataset.assembly == assembly
assert dataset.source == source
assert np.all(dataset.valid_chrs == ['chr7'])
assert np.all(dataset.test_chrs == ['chr8','chr9'])
assert(len(dataset.targetmap)==5)
assert(len(dataset.cellmap)==3)
def test_targets_SPI1_PAX5(self):
# https://github.com/YosefLab/epitome/issues/22
with warnings.catch_warnings(record=True) as warning_list:
warnings.simplefilter('always')
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = ['DNase','SPI1', 'PAX5'],min_cells_per_target=2, min_targets_per_cell=2)
self.assertTrue(len(warning_list) == 1) # one for SPI1 and PAX5
self.assertTrue(all(item.category == UserWarning for item in warning_list))
def test_specify_assays(self):
# test for https://github.com/YosefLab/epitome/issues/23
# should add DNase to eligible assays
eligible_targets = ['CTCF', 'RAD21', 'CEBPB']
dataset = EpitomeDataset(targets=eligible_targets)
model = EpitomeModel(dataset)
assert (len(model.dataset.targetmap) == 4)
def test_gets_correct_matrix_indices(self):
eligible_cells = ['IMR-90','H1','H9']
eligible_targets = ['DNase','H4K8ac']
matrix, cellmap, targetmap = EpitomeDataset.get_assays(
targets = eligible_targets,
cells = eligible_cells, min_cells_per_target = 3, min_targets_per_cell = 1)
self.assertTrue(matrix[cellmap['IMR-90']][targetmap['H4K8ac']]==0) # data for first row
def test_model_similarity_assays(self):
# should train a model without using DNAse
eligible_targets = ['CTCF', 'RAD21', 'CEBPB']
dataset = EpitomeDataset(targets=eligible_targets,
similarity_targets=['H3K27ac'])
model = EpitomeModel(dataset)
assert (len(model.dataset.targetmap) == 4)
def test_all_keys(self):
data = h5py.File(os.path.join(self.dataset.data_dir, EPITOME_H5_FILE), 'r')
keys = sorted(set(EpitomeDataset.all_keys(data)))
# bin size should be positive
self.assertTrue(data['columns']['binSize'][:][0] > 0)
data.close()
self.assertTrue(np.all([i in REQUIRED_KEYS for i in keys]))
def test_generator_sparse_data():
eligible_cells = ['K562', 'HepG2', 'H1', 'A549', 'HeLa-S3']
eligible_targets = ['DNase', 'CTCF', 'RAD21', 'LARP7']
dataset = EpitomeDataset(targets=eligible_targets,
cells=eligible_cells,
min_cells_per_target=1,
min_targets_per_cell=1)
label_cell_types = ['HepG2']
eligible_cells.remove(label_cell_types[0])
results = list(
load_data(dataset.get_data(Dataset.TRAIN),
label_cell_types,
eligible_cells,
dataset.matrix,
dataset.targetmap,
dataset.cellmap,
radii=[],
mode=Dataset.VALID,
return_feature_names=True,
indices=np.arange(0, 10))())
# get first features
features = results[0][0]
# get labels
labels = results[0][1]
# all cell types but K562 are missing LARP7 data
assert (len(features[0]) == len(eligible_cells) * len(eligible_targets) -
3)
# make sure mask is masking out LARP7 for HepG2
assert (np.all(features[-1] == [1., 0., 1.]))
# make sure first label cell is not the test cell K562
assert (labels[-2][0] == 'lbl_HepG2_RAD21')
assert (labels[-2][1] == 'lbl_HepG2_LARP7')
assert (labels[-2][2] == 'lbl_HepG2_CTCF')
def test_get_assays_single_target(self):
TF = ['DNase', 'JUND']
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = TF,
min_cells_per_target = 2,
min_targets_per_cell = 2)
targets = list(targetmap)
# Make sure only JUND and DNase are in list of targets
self.assertTrue(len(targets) == 2)
for t in TF:
self.assertTrue(t in targets)
def test_get_assays_without_DNase(self):
TF = 'JUND'
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = TF,
similarity_targets = ['H3K27ac'],
min_cells_per_target = 2,
min_targets_per_cell = 1)
targets = list(targetmap)
# Make sure only JUND and is in list of assays
self.assertTrue(len(targets) == 2)
self.assertTrue(TF in targets)
self.assertTrue('H3K27ac' in targets)
def test_generator_radius():
eligible_cells = ['K562', 'HepG2', 'H1', 'A549', 'HeLa-S3']
eligible_targets = ['DNase', 'CTCF', 'RAD21']
dataset = EpitomeDataset(targets=eligible_targets, cells=eligible_cells)
label_cell_types = ['K562']
eligible_cells.remove(label_cell_types[0])
radii = [1, 10]
results = load_data(dataset.get_data(Dataset.TRAIN), ['K562'],
eligible_cells,
dataset.matrix,
dataset.targetmap,
dataset.cellmap,
radii=radii,
mode=Dataset.VALID,
indices=np.arange(0, 10))()
li_results = list(results)
# length should include eligible targets and 2* radius for pos and agreement
assert (len(li_results[0][0]) == len(eligible_cells) *
(len(eligible_targets) + len(radii) * 2))
def test_correct_weights(self):
# make sure that assays with position 0 in matrix were cut off
# For example, TCFL2 in Panc1 has position 0 when loaded. It
# was previously being masked in the generator, even though the data was present.
ds = EpitomeDataset(targets=['TCF7L2'],
cells=['Panc1', 'MCF-7', 'K562'])
# make sure you are getting position 0
# this is where the bug was
assert np.where(ds.matrix == -1)[0].shape[0] == 0
model = EpitomeModel(ds)
model.train(1)
results = model.test(1000, calculate_metrics=True)
assert np.where(results['weights'] == 0)[0].shape[0] == 0
def test_reserve_validation_indices(self):
# new dataset because we are modifying it
dataset = EpitomeDataset()
self.assertTrue(dataset.get_data(Dataset.TRAIN).shape == (746, 1800))
self.assertTrue(dataset.get_data(Dataset.TRAIN_VALID).shape == (746,0))
old_indices = dataset.indices[Dataset.TRAIN]
self.assertTrue(dataset.indices[Dataset.TRAIN].shape[0] == 1800)
dataset.set_train_validation_indices("chr1")
self.assertTrue(dataset.get_data(Dataset.TRAIN).shape == (746, 1700))
self.assertTrue(dataset.get_data(Dataset.TRAIN_VALID).shape == (746, 100))
# check indices
self.assertTrue(dataset.indices[Dataset.TRAIN].shape[0] == 1700)
self.assertTrue(dataset.indices[Dataset.TRAIN_VALID].shape[0] == 100)
self.assertTrue(dataset.indices[Dataset.TRAIN_VALID][0] == 0) # first chr
self.assertTrue(dataset.indices[Dataset.TRAIN][0] == 100) # start of chr2
joined_indices = np.concatenate([dataset.indices[Dataset.TRAIN_VALID], dataset.indices[Dataset.TRAIN]])
joined_indices.sort()
self.assertTrue(len(np.setdiff1d(joined_indices, old_indices)) == 0 and len(np.setdiff1d(old_indices, joined_indices)) == 0)
class DatasetTest(EpitomeTestCase):
def __init__(self, *args, **kwargs):
super(DatasetTest, self).__init__(*args, **kwargs)
self.dataset = EpitomeDataset()
def test_user_data_path(self):
# user data path should be able to be explicitly set
datapath = GET_DATA_PATH()
self.assertTrue(datapath == os.environ["EPITOME_DATA_PATH"])
def test_save(self):
out_path = self.tmpFile()
print(out_path)
# generate 100 records for each chromosome
PER_CHR_RECORDS=10
chroms = []
newStarts = []
binSize = 200
for i in range(1,22): # for all 22 chrs
chrom = 'chr' + str(i)
for j in range(1,PER_CHR_RECORDS+1):
chroms.append(chrom)
newStarts.append(j * binSize)
regions_df = pd.DataFrame({'Chromosome': chroms, 'Start': newStarts})
targets = ['DNase', 'CTCF', 'Rad21', 'SCM3', 'IRF3']
cellTypes = ['K562'] * len(targets) + ['A549'] * len(targets) + ['GM12878'] * len(targets)
targets = targets * 3
row_df = pd.DataFrame({'cellType': cellTypes, 'target': targets})
# generate random data of 0/1s
np.random.seed(0)
rand_data = np.random.randint(2, size=(len(row_df), len(regions_df))).astype('i1')
source = 'rand_source'
assembly = 'hg19'
EpitomeDataset.save(out_path,
rand_data,
row_df,
regions_df,
binSize,
assembly,
source,
valid_chrs = ['chr7'],
test_chrs = ['chr8','chr9'])
# load back in dataset and make sure it checks out
dataset = EpitomeDataset(data_dir = out_path)
assert dataset.assembly == assembly
assert dataset.source == source
assert np.all(dataset.valid_chrs == ['chr7'])
assert np.all(dataset.test_chrs == ['chr8','chr9'])
assert(len(dataset.targetmap)==5)
assert(len(dataset.cellmap)==3)
def test_gets_correct_matrix_indices(self):
eligible_cells = ['IMR-90','H1','H9']
eligible_targets = ['DNase','H4K8ac']
matrix, cellmap, targetmap = EpitomeDataset.get_assays(
targets = eligible_targets,
cells = eligible_cells, min_cells_per_target = 3, min_targets_per_cell = 1)
self.assertTrue(matrix[cellmap['IMR-90']][targetmap['H4K8ac']]==0) # data for first row
def test_get_assays_single_target(self):
TF = ['DNase', 'JUND']
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = TF,
min_cells_per_target = 2,
min_targets_per_cell = 2)
targets = list(targetmap)
# Make sure only JUND and DNase are in list of targets
self.assertTrue(len(targets) == 2)
for t in TF:
self.assertTrue(t in targets)
def test_get_targets_without_DNase(self):
TF = 'JUND'
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = TF,
similarity_targets = ['H3K27ac'],
min_cells_per_target = 2,
min_targets_per_cell = 1)
targets = list(targetmap)
# Make sure only JUND and is in list of targets
self.assertTrue(len(targets) == 2)
self.assertTrue(TF in targets)
self.assertTrue('H3K27ac' in targets)
def test_list_targets(self):
targets =self.dataset.list_targets()
self.assertTrue(len(targets) == len(self.dataset.targetmap))
def test_get_assays_without_DNase(self):
TF = 'JUND'
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = TF,
similarity_targets = ['H3K27ac'],
min_cells_per_target = 2,
min_targets_per_cell = 1)
targets = list(targetmap)
# Make sure only JUND and is in list of assays
self.assertTrue(len(targets) == 2)
self.assertTrue(TF in targets)
self.assertTrue('H3K27ac' in targets)
def test_targets_SPI1_PAX5(self):
# https://github.com/YosefLab/epitome/issues/22
with warnings.catch_warnings(record=True) as warning_list:
warnings.simplefilter('always')
matrix, cellmap, targetmap = EpitomeDataset.get_assays(targets = ['DNase','SPI1', 'PAX5'],min_cells_per_target=2, min_targets_per_cell=2)
self.assertTrue(len(warning_list) == 1) # one for SPI1 and PAX5
self.assertTrue(all(item.category == UserWarning for item in warning_list))
def test_get_data(self):
train_data = self.dataset.get_data(Dataset.TRAIN)
self.assertTrue(train_data.shape[0] == np.where(self.dataset.matrix!= -1)[0].shape[0])
self.assertTrue(train_data.shape[1] == 1800)
valid_data = self.dataset.get_data(Dataset.VALID)
self.assertTrue(valid_data.shape[1] == 100)
test_data = self.dataset.get_data(Dataset.TEST)
self.assertTrue(test_data.shape[1] == 200)
all_data = self.dataset.get_data(Dataset.ALL)
self.assertTrue(all_data.shape[1] == 2100)
# Make sure you are getting the right data in the right order
alldata_filtered = self.dataset.get_data(Dataset.ALL)
dataset = h5py.File(os.path.join(self.dataset.data_dir, EPITOME_H5_FILE), 'r')
alldata= dataset['data'][:]
dataset.close()
self.assertTrue(np.all(alldata[self.dataset.full_matrix[:, self.dataset.targetmap['DNase']],:] == alldata_filtered[self.dataset.matrix[:, self.dataset.targetmap['DNase']],:]))
def test_order_by_similarity(self):
cell = list(self.dataset.cellmap)[0]
mode = Dataset.VALID
sim = self.dataset.order_by_similarity(cell, mode, compare_target = 'DNase')
self.assertTrue(len(sim) == len(self.dataset.cellmap))
self.assertTrue(sim[0] == cell)
def test_all_keys(self):
data = h5py.File(os.path.join(self.dataset.data_dir, EPITOME_H5_FILE), 'r')
keys = sorted(set(EpitomeDataset.all_keys(data)))
# bin size should be positive
self.assertTrue(data['columns']['binSize'][:][0] > 0)
data.close()
self.assertTrue(np.all([i in REQUIRED_KEYS for i in keys]))
def test_reserve_validation_indices(self):
# new dataset because we are modifying it
dataset = EpitomeDataset()
self.assertTrue(dataset.get_data(Dataset.TRAIN).shape == (746, 1800))
self.assertTrue(dataset.get_data(Dataset.TRAIN_VALID).shape == (746,0))
old_indices = dataset.indices[Dataset.TRAIN]
self.assertTrue(dataset.indices[Dataset.TRAIN].shape[0] == 1800)
dataset.set_train_validation_indices("chr1")
self.assertTrue(dataset.get_data(Dataset.TRAIN).shape == (746, 1700))
self.assertTrue(dataset.get_data(Dataset.TRAIN_VALID).shape == (746, 100))
# check indices
self.assertTrue(dataset.indices[Dataset.TRAIN].shape[0] == 1700)
self.assertTrue(dataset.indices[Dataset.TRAIN_VALID].shape[0] == 100)
self.assertTrue(dataset.indices[Dataset.TRAIN_VALID][0] == 0) # first chr
self.assertTrue(dataset.indices[Dataset.TRAIN][0] == 100) # start of chr2
joined_indices = np.concatenate([dataset.indices[Dataset.TRAIN_VALID], dataset.indices[Dataset.TRAIN]])
joined_indices.sort()
self.assertTrue(len(np.setdiff1d(joined_indices, old_indices)) == 0 and len(np.setdiff1d(old_indices, joined_indices)) == 0)
def __init__(self, *args, **kwargs):
super(DatasetTest, self).__init__(*args, **kwargs)
self.dataset = EpitomeDataset()
