def main():
usage = 'solo'
parser = ArgumentParser(usage,
formatter_class=ArgumentDefaultsHelpFormatter)
parser.add_argument(dest='model_json_file',
help='json file to pass VAE parameters')
parser.add_argument(
dest='data_path',
help=
'path to h5ad, loom or 10x directory containing cell by genes counts')
parser.add_argument('-d',
dest='doublet_depth',
default=2.,
type=float,
help='Depth multiplier for a doublet relative to the \
average of its constituents')
parser.add_argument('-g',
dest='gpu',
default=True,
action='store_true',
help='Run on GPU')
parser.add_argument('-a',
dest='anndata_output',
default=False,
action='store_true',
help='output modified anndata object with solo scores \
Only works for anndata')
parser.add_argument('-o', dest='out_dir', default='solo_out')
parser.add_argument('-r',
dest='doublet_ratio',
default=2.,
type=float,
help='Ratio of doublets to true \
cells')
parser.add_argument('-s',
dest='seed',
default=None,
help='Path to previous solo output \
directory. Seed VAE models with previously \
trained solo model. Directory structure is assumed to \
be the same as solo output directory structure. \
should at least have a vae.pt a pickled object of \
vae weights and a latent.npy an np.ndarray of the \
latents of your cells.')
parser.add_argument('-k',
dest='known_doublets',
help='Experimentally defined doublets tsv file. \
Should be a single column of True/False. True \
indicates the cell is a doublet. No header.',
type=str)
parser.add_argument('-t',
dest='doublet_type',
help='Please enter \
multinomial, average, or sum',
default='multinomial',
choices=['multinomial', 'average', 'sum'])
parser.add_argument('-e',
dest='expected_number_of_doublets',
help='Experimentally expected number of doublets',
type=int,
default=None)
parser.add_argument('-p',
dest='plot',
default=False,
action='store_true',
help='Plot outputs for solo')
parser.add_argument('-l',
dest='normal_logging',
default=False,
action='store_true',
help='Logging level set to normal (aka not debug)')
parser.add_argument('--random_size',
dest='randomize_doublet_size',
default=False,
action='store_true',
help='Sample depth multipliers from Unif(1, \
DoubletDepth) \
to provide a diversity of possible doublet depths.')
args = parser.parse_args()
if not args.normal_logging:
scvi._settings.set_verbosity(10)
model_json_file = args.model_json_file
data_path = args.data_path
if args.gpu and not torch.cuda.is_available():
args.gpu = torch.cuda.is_available()
print('Cuda is not available, switching to cpu running!')
if not os.path.isdir(args.out_dir):
os.mkdir(args.out_dir)
##################################################
# data
# read loom/anndata
data_ext = os.path.splitext(data_path)[-1]
if data_ext == '.loom':
scvi_data = LoomDataset(data_path)
elif data_ext == '.h5ad':
adata = anndata.read(data_path)
if issparse(adata.X):
adata.X = adata.X.todense()
scvi_data = AnnDatasetFromAnnData(adata)
elif os.path.isdir(data_path):
scvi_data = Dataset10X(save_path=data_path,
measurement_names_column=1,
dense=True)
cell_umi_depth = scvi_data.X.sum(axis=1)
fifth, ninetyfifth = np.percentile(cell_umi_depth, [5, 95])
min_cell_umi_depth = np.min(cell_umi_depth)
max_cell_umi_depth = np.max(cell_umi_depth)
if fifth * 10 < ninetyfifth:
print("""WARNING YOUR DATA HAS A WIDE RANGE OF CELL DEPTHS.
PLEASE MANUALLY REVIEW YOUR DATA""")
print(
f"Min cell depth: {min_cell_umi_depth}, Max cell depth: {max_cell_umi_depth}"
)
else:
msg = f'{data_path} is not a recognized format.\n'
msg += 'must be one of {h5ad, loom, 10x directory}'
raise TypeError(msg)
num_cells, num_genes = scvi_data.X.shape
if args.known_doublets is not None:
print('Removing known doublets for in silico doublet generation')
print('Make sure known doublets are in the same order as your data')
known_doublets = np.loadtxt(args.known_doublets, dtype=str) == 'True'
assert len(known_doublets) == scvi_data.X.shape[0]
known_doublet_data = make_gene_expression_dataset(
scvi_data.X[known_doublets], scvi_data.gene_names)
known_doublet_data.labels = np.ones(known_doublet_data.X.shape[0])
singlet_scvi_data = make_gene_expression_dataset(
scvi_data.X[~known_doublets], scvi_data.gene_names)
singlet_num_cells, _ = singlet_scvi_data.X.shape
else:
known_doublet_data = None
singlet_num_cells = num_cells
known_doublets = np.zeros(num_cells, dtype=bool)
singlet_scvi_data = scvi_data
singlet_scvi_data.labels = np.zeros(singlet_scvi_data.X.shape[0])
scvi_data.labels = known_doublets.astype(int)
##################################################
# parameters
# check for parameters
if not os.path.exists(model_json_file):
raise FileNotFoundError(f'{model_json_file} does not exist.')
# read parameters
with open(model_json_file, 'r') as model_json_open:
params = json.load(model_json_open)
# set VAE params
vae_params = {}
for par in [
'n_hidden', 'n_latent', 'n_layers', 'dropout_rate', 'ignore_batch'
]:
if par in params:
vae_params[par] = params[par]
vae_params['n_batch'] = 0 if params.get('ignore_batch',
False) else scvi_data.n_batches
# training parameters
batch_size = params.get('batch_size', 128)
valid_pct = params.get('valid_pct', 0.1)
learning_rate = params.get('learning_rate', 1e-3)
stopping_params = {'patience': params.get('patience', 10), 'threshold': 0}
# protect against single example batch
while num_cells % batch_size == 1:
batch_size = int(np.round(1.25 * batch_size))
print('Increasing batch_size to %d to avoid single example batch.' %
batch_size)
##################################################
# VAE
vae = VAE(n_input=singlet_scvi_data.nb_genes,
n_labels=2,
reconstruction_loss='nb',
log_variational=True,
**vae_params)
if args.seed:
if args.gpu:
device = torch.device('cuda')
vae.load_state_dict(torch.load(os.path.join(args.seed, 'vae.pt')))
vae.to(device)
else:
map_loc = 'cpu'
vae.load_state_dict(
torch.load(os.path.join(args.seed, 'vae.pt'),
map_location=map_loc))
# save latent representation
utrainer = \
UnsupervisedTrainer(vae, singlet_scvi_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['reconstruction_error'],
use_cuda=args.gpu,
early_stopping_kwargs=stopping_params,
batch_size=batch_size)
full_posterior = utrainer.create_posterior(utrainer.model,
singlet_scvi_data,
indices=np.arange(
len(singlet_scvi_data)))
latent, _, _ = full_posterior.sequential(batch_size).get_latent()
np.save(os.path.join(args.out_dir, 'latent.npy'),
latent.astype('float32'))
else:
stopping_params['early_stopping_metric'] = 'reconstruction_error'
stopping_params['save_best_state_metric'] = 'reconstruction_error'
# initialize unsupervised trainer
utrainer = \
UnsupervisedTrainer(vae, singlet_scvi_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['reconstruction_error'],
use_cuda=args.gpu,
early_stopping_kwargs=stopping_params,
batch_size=batch_size)
utrainer.history['reconstruction_error_test_set'].append(0)
# initial epoch
utrainer.train(n_epochs=2000, lr=learning_rate)
# drop learning rate and continue
utrainer.early_stopping.wait = 0
utrainer.train(n_epochs=500, lr=0.5 * learning_rate)
# save VAE
torch.save(vae.state_dict(), os.path.join(args.out_dir, 'vae.pt'))
# save latent representation
full_posterior = utrainer.create_posterior(utrainer.model,
singlet_scvi_data,
indices=np.arange(
len(singlet_scvi_data)))
latent, _, _ = full_posterior.sequential(batch_size).get_latent()
np.save(os.path.join(args.out_dir, 'latent.npy'),
latent.astype('float32'))
##################################################
# simulate doublets
non_zero_indexes = np.where(singlet_scvi_data.X > 0)
cells = non_zero_indexes[0]
genes = non_zero_indexes[1]
cells_ids = defaultdict(list)
for cell_id, gene in zip(cells, genes):
cells_ids[cell_id].append(gene)
# choose doublets function type
if args.doublet_type == 'average':
doublet_function = create_average_doublet
elif args.doublet_type == 'sum':
doublet_function = create_summed_doublet
else:
doublet_function = create_multinomial_doublet
cell_depths = singlet_scvi_data.X.sum(axis=1)
num_doublets = int(args.doublet_ratio * singlet_num_cells)
if known_doublet_data is not None:
num_doublets -= known_doublet_data.X.shape[0]
# make sure we are making a non negative amount of doublets
assert num_doublets >= 0
in_silico_doublets = np.zeros((num_doublets, num_genes), dtype='float32')
# for desired # doublets
for di in range(num_doublets):
# sample two cells
i, j = np.random.choice(singlet_num_cells, size=2)
# generate doublets
in_silico_doublets[di, :] = \
doublet_function(singlet_scvi_data.X, i, j,
doublet_depth=args.doublet_depth,
cell_depths=cell_depths, cells_ids=cells_ids,
randomize_doublet_size=args.randomize_doublet_size)
# merge datasets
# we can maybe up sample the known doublets
# concatentate
classifier_data = GeneExpressionDataset()
classifier_data.populate_from_data(
X=np.vstack([scvi_data.X, in_silico_doublets]),
labels=np.hstack(
[np.ravel(scvi_data.labels),
np.ones(in_silico_doublets.shape[0])]),
remap_attributes=False)
assert (len(np.unique(classifier_data.labels.flatten())) == 2)
##################################################
# classifier
# model
classifier = Classifier(n_input=(vae.n_latent + 1),
n_hidden=params['cl_hidden'],
n_layers=params['cl_layers'],
n_labels=2,
dropout_rate=params['dropout_rate'])
# trainer
stopping_params['early_stopping_metric'] = 'accuracy'
stopping_params['save_best_state_metric'] = 'accuracy'
strainer = ClassifierTrainer(classifier,
classifier_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['accuracy'],
use_cuda=args.gpu,
sampling_model=vae,
sampling_zl=True,
early_stopping_kwargs=stopping_params,
batch_size=batch_size)
# initial
strainer.train(n_epochs=1000, lr=learning_rate)
# drop learning rate and continue
strainer.early_stopping.wait = 0
strainer.train(n_epochs=300, lr=0.1 * learning_rate)
torch.save(classifier.state_dict(),
os.path.join(args.out_dir, 'classifier.pt'))
##################################################
# post-processing
# use logits for predictions for better results
logits_classifier = Classifier(n_input=(vae.n_latent + 1),
n_hidden=params['cl_hidden'],
n_layers=params['cl_layers'],
n_labels=2,
dropout_rate=params['dropout_rate'],
logits=True)
logits_classifier.load_state_dict(classifier.state_dict())
# using logits leads to better performance in for ranking
logits_strainer = ClassifierTrainer(logits_classifier,
classifier_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['accuracy'],
use_cuda=args.gpu,
sampling_model=vae,
sampling_zl=True,
early_stopping_kwargs=stopping_params,
batch_size=batch_size)
# models evaluation mode
vae.eval()
classifier.eval()
logits_classifier.eval()
print('Train accuracy: %.4f' % strainer.train_set.accuracy())
print('Test accuracy: %.4f' % strainer.test_set.accuracy())
# compute predictions manually
# output logits
train_y, train_score = strainer.train_set.compute_predictions(soft=True)
test_y, test_score = strainer.test_set.compute_predictions(soft=True)
# train_y == true label
# train_score[:, 0] == singlet score; train_score[:, 1] == doublet score
train_score = train_score[:, 1]
train_y = train_y.astype('bool')
test_score = test_score[:, 1]
test_y = test_y.astype('bool')
train_auroc = roc_auc_score(train_y, train_score)
test_auroc = roc_auc_score(test_y, test_score)
print('Train AUROC: %.4f' % train_auroc)
print('Test AUROC: %.4f' % test_auroc)
train_fpr, train_tpr, train_t = roc_curve(train_y, train_score)
test_fpr, test_tpr, test_t = roc_curve(test_y, test_score)
train_t = np.minimum(train_t, 1 + 1e-9)
test_t = np.minimum(test_t, 1 + 1e-9)
train_acc = np.zeros(len(train_t))
for i in range(len(train_t)):
train_acc[i] = np.mean(train_y == (train_score > train_t[i]))
test_acc = np.zeros(len(test_t))
for i in range(len(test_t)):
test_acc[i] = np.mean(test_y == (test_score > test_t[i]))
# write predictions
# softmax predictions
order_y, order_score = strainer.compute_predictions(soft=True)
_, order_pred = strainer.compute_predictions()
doublet_score = order_score[:, 1]
np.save(os.path.join(args.out_dir, 'no_updates_softmax_scores.npy'),
doublet_score[:num_cells])
np.save(os.path.join(args.out_dir, 'no_updates_softmax_scores_sim.npy'),
doublet_score[num_cells:])
# logit predictions
logit_y, logit_score = logits_strainer.compute_predictions(soft=True)
logit_doublet_score = logit_score[:, 1]
np.save(os.path.join(args.out_dir, 'logit_scores.npy'),
logit_doublet_score[:num_cells])
np.save(os.path.join(args.out_dir, 'logit_scores_sim.npy'),
logit_doublet_score[num_cells:])
# update threshold as a function of Solo's estimate of the number of
# doublets
# essentially a log odds update
# TODO put in a function
diff = np.inf
counter_update = 0
solo_scores = doublet_score[:num_cells]
logit_scores = logit_doublet_score[:num_cells]
d_s = (args.doublet_ratio / (args.doublet_ratio + 1))
while (diff > .01) | (counter_update < 5):
# calculate log odss calibration for logits
d_o = np.mean(solo_scores)
c = np.log(d_o / (1 - d_o)) - np.log(d_s / (1 - d_s))
# update soloe scores
solo_scores = 1 / (1 + np.exp(-(logit_scores + c)))
# update while conditions
diff = np.abs(d_o - np.mean(solo_scores))
counter_update += 1
np.save(os.path.join(args.out_dir, 'softmax_scores.npy'), solo_scores)
if args.expected_number_of_doublets is not None:
k = len(solo_scores) - args.expected_number_of_doublets
if args.expected_number_of_doublets / len(solo_scores) > .5:
print('''Make sure you actually expect more than half your cells
to be doublets. If not change your
-e parameter value''')
assert k > 0
idx = np.argpartition(solo_scores, k)
threshold = np.max(solo_scores[idx[:k]])
is_solo_doublet = solo_scores > threshold
else:
is_solo_doublet = solo_scores > .5
is_doublet = known_doublets
new_doublets_idx = np.where(~(is_doublet) & is_solo_doublet[:num_cells])[0]
is_doublet[new_doublets_idx] = True
np.save(os.path.join(args.out_dir, 'is_doublet.npy'),
is_doublet[:num_cells])
np.save(os.path.join(args.out_dir, 'is_doublet_sim.npy'),
is_doublet[num_cells:])
np.save(os.path.join(args.out_dir, 'preds.npy'), order_pred[:num_cells])
np.save(os.path.join(args.out_dir, 'preds_sim.npy'),
order_pred[num_cells:])
smoothed_preds = knn_smooth_pred_class(X=latent,
pred_class=is_doublet[:num_cells])
np.save(os.path.join(args.out_dir, 'smoothed_preds.npy'), smoothed_preds)
if args.anndata_output and data_ext == '.h5ad':
adata.obs['is_doublet'] = is_doublet[:num_cells]
adata.obs['logit_scores'] = logit_doublet_score[:num_cells]
adata.obs['softmax_scores'] = doublet_score[:num_cells]
adata.write(os.path.join(args.out_dir, "soloed.h5ad"))
if args.plot:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import seaborn as sns
# plot ROC
plt.figure()
plt.plot(train_fpr, train_tpr, label='Train')
plt.plot(test_fpr, test_tpr, label='Test')
plt.gca().set_xlabel('False positive rate')
plt.gca().set_ylabel('True positive rate')
plt.legend()
plt.savefig(os.path.join(args.out_dir, 'roc.pdf'))
plt.close()
# plot accuracy
plt.figure()
plt.plot(train_t, train_acc, label='Train')
plt.plot(test_t, test_acc, label='Test')
plt.axvline(0.5, color='black', linestyle='--')
plt.gca().set_xlabel('Threshold')
plt.gca().set_ylabel('Accuracy')
plt.legend()
plt.savefig(os.path.join(args.out_dir, 'accuracy.pdf'))
plt.close()
# plot distributions
plt.figure()
sns.distplot(test_score[test_y], label='Simulated')
sns.distplot(test_score[~test_y], label='Observed')
plt.legend()
plt.savefig(os.path.join(args.out_dir, 'train_v_test_dist.pdf'))
plt.close()
plt.figure()
sns.distplot(doublet_score[:num_cells], label='Observed')
plt.legend()
plt.savefig(os.path.join(args.out_dir, 'real_cells_dist.pdf'))
plt.close()
scvi_umap = umap.UMAP(n_neighbors=16).fit_transform(latent)
fig, ax = plt.subplots(1, 1, figsize=(10, 10))
ax.scatter(scvi_umap[:, 0],
scvi_umap[:, 1],
c=doublet_score[:num_cells],
s=8,
cmap="GnBu")
ax.set_xlabel("UMAP 1")
ax.set_ylabel("UMAP 2")
ax.set_xticks([], [])
ax.set_yticks([], [])
fig.savefig(os.path.join(args.out_dir, 'umap_solo_scores.pdf'))
def solo(X,
gene_names,
doublet_depth=2.0,
gpu=False,
out_dir='solo_out',
doublet_ratio=2.0,
seed=None,
known_doublets=None,
doublet_type='multinomial',
expected_number_of_doublets=None,
plot=False,
normal_logging=False,
n_hidden=128,
n_latent=16,
cl_hidden=64,
cl_layers=1,
dropout_rate=0.1,
learning_rate=0.001,
valid_pct=0.1):
from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter
import json
import os
import shutil
import anndata
import numpy as np
from anndata import AnnData
from sklearn.metrics import roc_auc_score, roc_curve
from scipy.sparse import issparse
from collections import defaultdict
import scvi
from scvi.dataset import AnnDatasetFromAnnData, LoomDataset, GeneExpressionDataset
from scvi.models import Classifier, VAE
from scvi.inference import UnsupervisedTrainer, ClassifierTrainer
import torch
from solo.utils import create_average_doublet, create_summed_doublet, create_multinomial_doublet, make_gene_expression_dataset
if not normal_logging:
scvi._settings.set_verbosity(10)
if gpu and not torch.cuda.is_available():
gpu = torch.cuda.is_available()
print('Cuda is not available, switching to cpu running!')
# if not os.path.isdir(out_dir):
# os.mkdir(out_dir)
# data_ext = os.path.splitext(data_file)[-1]
# if data_ext == '.loom':
# scvi_data = LoomDataset(data_file)
# elif data_ext == '.h5ad':
# scvi_data = AnnDatasetFromAnnData(anndata.read(data_file))
# else:
# msg = f'{data_ext} is not a recognized format.\n'
# msg += 'must be one of {h5ad, loom}'
# raise TypeError(msg)
# if issparse(scvi_data.X):
# scvi_data.X = scvi_data.X.todense()
scvi_data = make_gene_expression_dataset(X, gene_names)
num_cells, num_genes = scvi_data.X.shape
if known_doublets is not None:
print('Removing known doublets for in silico doublet generation')
print('Make sure known doublets are in the same order as your data')
known_doublets = np.loadtxt(known_doublets, dtype=str) == 'True'
assert len(known_doublets) == scvi_data.X.shape[0]
known_doublet_data = make_gene_expression_dataset(
scvi_data.X[known_doublets], scvi_data.gene_names)
known_doublet_data.labels = np.ones(known_doublet_data.X.shape[0])
singlet_scvi_data = make_gene_expression_dataset(
scvi_data.X[~known_doublets], scvi_data.gene_names)
singlet_num_cells, _ = singlet_scvi_data.X.shape
else:
known_doublet_data = None
singlet_num_cells = num_cells
known_doublets = np.zeros(num_cells, dtype=bool)
singlet_scvi_data = scvi_data
singlet_scvi_data.labels = np.zeros(singlet_scvi_data.X.shape[0])
scvi_data.labels = known_doublets.astype(int)
params = {
"n_hidden": n_hidden,
"n_latent": n_latent,
"cl_hidden": cl_hidden,
"cl_layers": cl_layers,
"dropout_rate": dropout_rate,
"learning_rate": learning_rate,
"valid_pct": valid_pct
}
# set VAE params
vae_params = {}
for par in [
'n_hidden', 'n_latent', 'n_layers', 'dropout_rate', 'ignore_batch'
]:
if par in params:
vae_params[par] = params[par]
vae_params['n_batch'] = 0 if params.get('ignore_batch',
False) else scvi_data.n_batches
# training parameters
valid_pct = params.get('valid_pct', 0.1)
learning_rate = params.get('learning_rate', 1e-3)
stopping_params = {'patience': params.get('patience', 10), 'threshold': 0}
##################################################
# VAE
vae = VAE(n_input=singlet_scvi_data.nb_genes,
n_labels=2,
reconstruction_loss='nb',
log_variational=True,
**vae_params)
if seed:
if gpu:
device = torch.device('cuda')
vae.load_state_dict(torch.load(os.path.join(seed, 'vae.pt')))
vae.to(device)
else:
map_loc = 'cpu'
vae.load_state_dict(
torch.load(os.path.join(seed, 'vae.pt'), map_location=map_loc))
# copy latent representation
latent_file = os.path.join(seed, 'latent.npy')
if os.path.isfile(latent_file):
shutil.copy(latent_file, os.path.join(out_dir, 'latent.npy'))
else:
stopping_params['early_stopping_metric'] = 'reconstruction_error'
stopping_params['save_best_state_metric'] = 'reconstruction_error'
# initialize unsupervised trainer
utrainer = \
UnsupervisedTrainer(vae, singlet_scvi_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['reconstruction_error'],
use_cuda=gpu,
early_stopping_kwargs=stopping_params)
utrainer.history['reconstruction_error_test_set'].append(0)
# initial epoch
utrainer.train(n_epochs=2000, lr=learning_rate)
# drop learning rate and continue
utrainer.early_stopping.wait = 0
utrainer.train(n_epochs=500, lr=0.5 * learning_rate)
# save VAE
torch.save(vae.state_dict(), os.path.join(out_dir, 'vae.pt'))
# save latent representation
full_posterior = utrainer.create_posterior(utrainer.model,
singlet_scvi_data,
indices=np.arange(
len(singlet_scvi_data)))
latent, _, _ = full_posterior.sequential().get_latent()
np.save(os.path.join(out_dir, 'latent.npy'), latent.astype('float32'))
##################################################
# simulate doublets
non_zero_indexes = np.where(singlet_scvi_data.X > 0)
cells = non_zero_indexes[0]
genes = non_zero_indexes[1]
cells_ids = defaultdict(list)
for cell_id, gene in zip(cells, genes):
cells_ids[cell_id].append(gene)
# choose doublets function type
if doublet_type == 'average':
doublet_function = create_average_doublet
elif doublet_type == 'sum':
doublet_function = create_summed_doublet
else:
doublet_function = create_multinomial_doublet
cell_depths = singlet_scvi_data.X.sum(axis=1)
num_doublets = int(doublet_ratio * singlet_num_cells)
if known_doublet_data is not None:
num_doublets -= known_doublet_data.X.shape[0]
# make sure we are making a non negative amount of doublets
assert num_doublets >= 0
in_silico_doublets = np.zeros((num_doublets, num_genes), dtype='float32')
# for desired # doublets
for di in range(num_doublets):
# sample two cells
i, j = np.random.choice(singlet_num_cells, size=2)
# generate doublets
in_silico_doublets[di, :] = \
doublet_function(singlet_scvi_data.X, i, j,
doublet_depth=doublet_depth,
cell_depths=cell_depths, cells_ids=cells_ids)
# merge datasets
# we can maybe up sample the known doublets
# concatentate
classifier_data = GeneExpressionDataset()
classifier_data.populate_from_data(
X=np.vstack([scvi_data.X, in_silico_doublets]),
labels=np.hstack(
[np.ravel(scvi_data.labels),
np.ones(in_silico_doublets.shape[0])]),
remap_attributes=False)
assert (len(np.unique(classifier_data.labels.flatten())) == 2)
##################################################
# classifier
# model
classifier = Classifier(n_input=(vae.n_latent + 1),
n_hidden=params['cl_hidden'],
n_layers=params['cl_layers'],
n_labels=2,
dropout_rate=params['dropout_rate'])
# trainer
stopping_params['early_stopping_metric'] = 'accuracy'
stopping_params['save_best_state_metric'] = 'accuracy'
strainer = ClassifierTrainer(classifier,
classifier_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['accuracy'],
use_cuda=gpu,
sampling_model=vae,
sampling_zl=True,
early_stopping_kwargs=stopping_params)
# initial
strainer.train(n_epochs=1000, lr=learning_rate)
# drop learning rate and continue
strainer.early_stopping.wait = 0
strainer.train(n_epochs=300, lr=0.1 * learning_rate)
torch.save(classifier.state_dict(), os.path.join(out_dir, 'classifier.pt'))
##################################################
# post-processing
# use logits for predictions for better results
logits_classifier = Classifier(n_input=(vae.n_latent + 1),
n_hidden=params['cl_hidden'],
n_layers=params['cl_layers'],
n_labels=2,
dropout_rate=params['dropout_rate'],
logits=True)
logits_classifier.load_state_dict(classifier.state_dict())
# using logits leads to better performance in for ranking
logits_strainer = ClassifierTrainer(logits_classifier,
classifier_data,
train_size=(1. - valid_pct),
frequency=2,
metrics_to_monitor=['accuracy'],
use_cuda=gpu,
sampling_model=vae,
sampling_zl=True,
early_stopping_kwargs=stopping_params)
# models evaluation mode
vae.eval()
classifier.eval()
logits_classifier.eval()
print('Train accuracy: %.4f' % strainer.train_set.accuracy())
print('Test accuracy: %.4f' % strainer.test_set.accuracy())
# compute predictions manually
# output logits
train_y, train_score = strainer.train_set.compute_predictions(soft=True)
test_y, test_score = strainer.test_set.compute_predictions(soft=True)
# train_y == true label
# train_score[:, 0] == singlet score; train_score[:, 1] == doublet score
train_score = train_score[:, 1]
train_y = train_y.astype('bool')
test_score = test_score[:, 1]
test_y = test_y.astype('bool')
train_auroc = roc_auc_score(train_y, train_score)
test_auroc = roc_auc_score(test_y, test_score)
print('Train AUROC: %.4f' % train_auroc)
print('Test AUROC: %.4f' % test_auroc)
train_fpr, train_tpr, train_t = roc_curve(train_y, train_score)
test_fpr, test_tpr, test_t = roc_curve(test_y, test_score)
train_t = np.minimum(train_t, 1 + 1e-9)
test_t = np.minimum(test_t, 1 + 1e-9)
train_acc = np.zeros(len(train_t))
for i in range(len(train_t)):
train_acc[i] = np.mean(train_y == (train_score > train_t[i]))
test_acc = np.zeros(len(test_t))
for i in range(len(test_t)):
test_acc[i] = np.mean(test_y == (test_score > test_t[i]))
# write predictions
# softmax predictions
order_y, order_score = strainer.compute_predictions(soft=True)
_, order_pred = strainer.compute_predictions()
doublet_score = order_score[:, 1]
np.save(os.path.join(out_dir, 'softmax_scores.npy'),
doublet_score[:num_cells])
np.save(os.path.join(out_dir, 'softmax_scores_sim.npy'),
doublet_score[num_cells:])
# logit predictions
logit_y, logit_score = logits_strainer.compute_predictions(soft=True)
logit_doublet_score = logit_score[:, 1]
np.save(os.path.join(out_dir, 'logit_scores.npy'),
logit_doublet_score[:num_cells])
np.save(os.path.join(out_dir, 'logit_scores_sim.npy'),
logit_doublet_score[num_cells:])
if expected_number_of_doublets is not None:
solo_scores = doublet_score[:num_cells]
k = len(solo_scores) - expected_number_of_doublets
if expected_number_of_doublets / len(solo_scores) > .5:
print(
'Make sure you actually expect more than half your cells to be doublets. If not change your -e parameter value'
)
assert k > 0
idx = np.argpartition(solo_scores, k)
threshold = np.max(solo_scores[idx[:k]])
is_solo_doublet = doublet_score > threshold
else:
is_solo_doublet = order_pred[:num_cells]
is_doublet = known_doublets
new_doublets_idx = np.where(~(is_doublet) & is_solo_doublet[:num_cells])[0]
is_doublet[new_doublets_idx] = True
np.save(os.path.join(out_dir, 'is_doublet.npy'), is_doublet[:num_cells])
np.save(os.path.join(out_dir, 'is_doublet_sim.npy'),
is_doublet[num_cells:])
np.save(os.path.join(out_dir, 'preds.npy'), order_pred[:num_cells])
np.save(os.path.join(out_dir, 'preds_sim.npy'), order_pred[num_cells:])
if plot:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import seaborn as sns
# plot ROC
plt.figure()
plt.plot(train_fpr, train_tpr, label='Train')
plt.plot(test_fpr, test_tpr, label='Test')
plt.gca().set_xlabel('False positive rate')
plt.gca().set_ylabel('True positive rate')
plt.legend()
plt.savefig(os.path.join(out_dir, 'roc.pdf'))
plt.close()
# plot accuracy
plt.figure()
plt.plot(train_t, train_acc, label='Train')
plt.plot(test_t, test_acc, label='Test')
plt.axvline(0.5, color='black', linestyle='--')
plt.gca().set_xlabel('Threshold')
plt.gca().set_ylabel('Accuracy')
plt.legend()
plt.savefig(os.path.join(out_dir, 'accuracy.pdf'))
plt.close()
# plot distributions
plt.figure()
sns.distplot(test_score[test_y], label='Simulated')
sns.distplot(test_score[~test_y], label='Observed')
plt.legend()
plt.savefig(os.path.join(out_dir, 'train_v_test_dist.pdf'))
plt.close()
plt.figure()
sns.distplot(doublet_score[:num_cells], label='Simulated')
plt.legend()
plt.savefig(os.path.join(out_dir, 'real_cells_dist.pdf'))
plt.close()
