def entropy_batch_mixing(self, verbose=False, **kwargs):
if self.gene_dataset.n_batches == 2:
latent, batch_indices, labels = self.get_latent()
sample = select_indices_evenly(2000, batch_indices)
be_score = entropy_batch_mixing(latent[sample, :],
batch_indices[sample], **kwargs)
if verbose:
print("Entropy batch mixing :", be_score)
return be_score
def eval_latent(batch_indices,
labels,
latent,
keys,
labelled_idx=None,
unlabelled_idx=None,
plotname=None,
plotting=False,
partial_only=True):
res_knn_partial = clustering_scores(latent, labels, 'knn', True,
labelled_idx, unlabelled_idx)
res_kmeans_partial = clustering_scores(latent, labels, 'KMeans', True,
labelled_idx, unlabelled_idx)
if partial_only == False:
res_knn = clustering_scores(np.asarray(latent), labels, 'knn')
res_kmeans = clustering_scores(np.asarray(latent), labels, 'KMeans')
# sample = select_indices_evenly(2000, batch_indices)
# batch_entropy = entropy_batch_mixing(latent[sample, :], batch_indices[sample])
# print("Entropy batch mixing :", batch_entropy)
if plotting == True and (os.path.isfile('../' + plotname + '.labels.pdf')
is False):
sample = select_indices_evenly(2000, batch_indices)
# sample = select_indices_evenly(2000, labels)
if plotname is not None:
colors = sns.color_palette('bright') + \
sns.color_palette('muted') + \
sns.color_palette('dark') + \
sns.color_palette('pastel') + \
sns.color_palette('colorblind')
latent_s = latent[sample, :]
label_s = labels[sample]
batch_s = batch_indices[sample]
if latent_s.shape[1] != 2:
latent_s = UMAP(spread=2).fit_transform(latent_s)
fig, ax = plt.subplots(figsize=(18, 18))
key_order = np.argsort(keys)
for i, k in enumerate(key_order):
ax.scatter(latent_s[label_s == k, 0],
latent_s[label_s == k, 1],
c=colors[i % 30],
label=keys[k],
edgecolors='none')
# ax.legend(bbox_to_anchor=(1.1, 0.5), borderaxespad=0, fontsize='x-large')
fig.patch.set_visible(False)
ax.axis('off')
fig.tight_layout()
plt.savefig('../' + plotname + '.labels.pdf')
fig, ax = plt.subplots(figsize=(18, 18))
# for i, x in enumerate(batch):
ax.scatter(latent_s[:, 0], latent_s[:, 1], c=batch_s, alpha=0.8)
# ax.legend(bbox_to_anchor=(1.1, 0.5), borderaxespad=0, fontsize='x-large')
plt.axis('off')
plt.savefig('../' + plotname + '.batchid.pdf')
if partial_only == False:
return res_knn, res_knn_partial, res_kmeans, res_kmeans_partial
else:
return 0, res_knn_partial, 0, res_kmeans_partial
def VAEstats(full):
ll = full.ll(verbose=True)
latent, batch_indices, labels = full.sequential().get_latent()
batch_indices = batch_indices.ravel()
if len(np.unique(batch_indices)) == 2:
sample = select_indices_evenly(
np.min(np.unique(batch_indices, return_counts=True)[1]),
batch_indices)
batch_entropy = entropy_batch_mixing(latent[sample, :],
batch_indices[sample])
else:
batch_entropy = -1
labels = labels.ravel()
stats = [ll, batch_entropy, -1, -1, np.arange(0, len(labels))]
return latent, batch_indices, labels, stats
def SCANVIstats(trainer_scanvi, gene_dataset):
full = trainer_scanvi.create_posterior(trainer_scanvi.model,
gene_dataset,
indices=np.arange(
len(gene_dataset)))
ll = full.ll(verbose=True)
latent, batch_indices, labels = full.sequential().get_latent()
batch_indices = batch_indices.ravel()
if len(np.unique(batch_indices)) == 2:
sample = select_indices_evenly(
np.min(np.unique(batch_indices, return_counts=True)[1]),
batch_indices)
batch_entropy = entropy_batch_mixing(latent[sample, :],
batch_indices[sample])
else:
batch_entropy = -1
labelled_idx = trainer_scanvi.labelled_set.indices
unlabelled_idx = trainer_scanvi.unlabelled_set.indices
trainer_scanvi.unlabelled_set = trainer_scanvi.create_posterior(
trainer_scanvi.model, gene_dataset, indices=unlabelled_idx)
acc = trainer_scanvi.unlabelled_set.accuracy()
stats = [ll, batch_entropy, acc, labelled_idx, unlabelled_idx]
return latent, batch_indices, labels, stats
"correlation between the cell-type composition of the subsampled dataset is %.3f"
% correlation)
sub_dataset = deepcopy(gene_dataset)
sub_dataset.update_cells(np.concatenate(cells))
vae = VAE(sub_dataset.nb_genes,
n_batch=sub_dataset.n_batches,
n_labels=sub_dataset.n_labels,
n_hidden=128,
dispersion='gene')
infer = VariationalInference(vae, sub_dataset, use_cuda=use_cuda)
infer.train(n_epochs=250)
latent, batch_indices, labels = infer.get_latent('sequential')
keys = sub_dataset.cell_types
batch_entropy = entropy_batch_mixing(latent, batch_indices)
print("Entropy batch mixing :", batch_entropy)
sample = select_indices_evenly(1000, labels)
res = knn_purity_avg(latent[sample, :],
labels[sample].astype('int'),
keys=keys,
acc=True)
print('average classification accuracy per cluster')
for x in res:
print(x)
knn_acc = np.mean([x[1] for x in res])
print("average KNN accuracy:", knn_acc)
res = clustering_scores(
np.asarray(latent)[sample, :], labels[sample], 'knn',
len(np.unique(labels[sample])))
for x in res:
print(x, res[x])
# latent, batch_indices,labels,keys = SEURAT.get_cca()
latent = np.genfromtxt('../macosko_regev.CCA.txt')
label = np.genfromtxt('../macosko_regev.CCA.label.txt',dtype='str')
keys = gene_dataset.cell_types
batch_indices = np.genfromtxt('../macosko_regev.CCA.batch.txt')
elif model_type == 'Combat':
COMBAT = COMBAT()
latent = COMBAT.combat_pca(gene_dataset)
latent = latent.T
batch_indices = np.concatenate(gene_dataset.batch_indices)
labels = np.concatenate(gene_dataset.labels)
keys = gene_dataset.cell_types
sample = select_indices_evenly(2000,batch_indices)
batch_entropy = entropy_batch_mixing(latent[sample, :], batch_indices[sample])
print("Entropy batch mixing :", batch_entropy)
sample = select_indices_evenly(1000,labels)
res = knn_purity_avg(
latent[sample, :], labels[sample],
keys=keys[np.unique(labels)], acc=True
)
print('average classification accuracy per cluster',np.mean([x[1] for x in res]))
for x in res:
print(x)
res = clustering_scores(np.asarray(latent)[sample,:],labels[sample],'knn',len(np.unique(labels[sample])))
def CompareModels(gene_dataset, dataset1, dataset2, plotname, models):
KNeighbors = np.concatenate(
[np.arange(10, 100, 10),
np.arange(100, 500, 50)])
K_int = np.concatenate([np.repeat(10, 10), np.repeat(50, 7)])
f = open('../' + plotname + '/' + models + '.res.txt', "w+")
f.write("model_type " + \
"knn_asw knn_nmi knn_ari knn_uca knn_wuca " + \
"p_knn_asw p_knn_nmi p_knn_ari p_knn_uca p_knn_wuca " + \
"p1_knn_asw p1_knn_nmi p1_knn_ari p1_knn_uca p1_knn_wuca " + \
"p2_knn_asw p2_knn_nmi p2_knn_ari p2_knn_uca p2_knn_wuca " + \
"kmeans_asw kmeans_nmi kmeans_ari kmeans_uca kmeans_wuca " + \
"p_kmeans_asw p_kmeans_nmi p_kmeans_ari p_kmeans_uca p_kmeans_wuca " + \
"p1_kmeans_asw p1_kmeans_nmi p1_kmeans_ari p1_kmeans_uca p1_kmeans_wuca " + \
"p2_kmeans_asw p2_kmeans_nmi p2_kmeans_ari p2_kmeans_uca p2_kmeans_wuca " + \
" ".join(['res_jaccard' + x for x in
np.concatenate([np.repeat(10, 10), np.repeat(50, 7)]).astype('str')]) + " " + \
'jaccard_score likelihood BE classifier_acc\n'
)
g = open('../' + plotname + '/' + models + '.percluster.res.txt', "w+")
g.write("model_type\tannotation\t" + "\t".join(gene_dataset.cell_types) +
"\n")
scanvi = SCANVI(gene_dataset.nb_genes, gene_dataset.n_batches,
gene_dataset.n_labels)
trainer_scanvi = SemiSupervisedTrainer(scanvi,
gene_dataset,
classification_ratio=1,
n_epochs_classifier=1,
lr_classification=5 * 1e-3)
labelled_idx = trainer_scanvi.labelled_set.indices
unlabelled_idx = trainer_scanvi.unlabelled_set.indices
if models == 'others':
latent1 = np.genfromtxt('../harmonization/Seurat_data/' + plotname +
'.1.CCA.txt')
latent2 = np.genfromtxt('../harmonization/Seurat_data/' + plotname +
'.2.CCA.txt')
for model_type in [
'scmap', 'readSeurat', 'coral', 'Combat', 'MNN', 'PCA'
]:
print(model_type)
if (model_type == 'scmap') or (model_type == 'coral'):
latent, batch_indices, labels, keys, stats = run_model(
model_type,
gene_dataset,
dataset1,
dataset2,
filename=plotname)
pred1 = latent
pred2 = stats
res1 = scmap_eval(pred1, labels[batch_indices == 1], labels)
res2 = scmap_eval(pred2, labels[batch_indices == 0], labels)
g.write("%s\t" % (model_type) + "p1\t" +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res1['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + "p2\t" +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res2['clusteracc']) + "\n"))
res = [-1] * 10 + \
[-1] + [res1[x] for x in ['nmi', 'ari', 'ca', 'weighted ca']] + \
[-1] + [res2[x] for x in ['nmi', 'ari', 'ca', 'weighted ca']] + \
[-1] * 41
f.write(model_type + (" %.4f" * 61 + "\n") % tuple(res))
else:
if model_type == 'readSeurat':
dataset1, dataset2, gene_dataset = SubsetGenes(
dataset1, dataset2, gene_dataset, plotname)
latent, batch_indices, labels, keys, stats = run_model(
model_type,
gene_dataset,
dataset1,
dataset2,
filename=plotname)
res_jaccard = [
KNNJaccardIndex(latent1, latent2, latent, batch_indices,
k)[0] for k in KNeighbors
]
res_jaccard_score = np.sum(res_jaccard * K_int)
res_knn, res_knn_partial, res_kmeans, res_kmeans_partial = \
eval_latent(batch_indices, labels, latent, keys,
labelled_idx, unlabelled_idx,
plotname=plotname + '.' + model_type, plotting=False, partial_only=False)
_, res_knn_partial1, _, res_kmeans_partial1 = \
eval_latent(batch_indices, labels, latent, keys,
batch_indices == 0, batch_indices == 1,
plotname=plotname + '.' + model_type, plotting=False)
_, res_knn_partial2, _, res_kmeans_partial2 = \
eval_latent(batch_indices, labels, latent, keys,
batch_indices == 1, batch_indices == 0,
plotname=plotname + '.' + model_type, plotting=False)
sample = select_indices_evenly(
np.min(np.unique(batch_indices, return_counts=True)[1]),
batch_indices)
batch_entropy = entropy_batch_mixing(latent[sample, :],
batch_indices[sample])
res = [res_knn[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_knn_partial[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_knn_partial1[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_knn_partial2[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_kmeans[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
[res_kmeans_partial[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
[res_kmeans_partial1[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
[res_kmeans_partial2[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
res_jaccard + \
[res_jaccard_score, -1, batch_entropy, -1]
f.write(model_type + (" %.4f" * 61 + "\n") % tuple(res))
g.write("%s\t" % (model_type) + 'all\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + 'p\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn_partial['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + 'p1\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn_partial1['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + 'p2\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn_partial2['clusteracc']) + "\n"))
elif (models == 'scvi') or (models == 'scvi_nb'):
dataset1, dataset2, gene_dataset = SubsetGenes(dataset1, dataset2,
gene_dataset, plotname)
if models == 'scvi_nb':
latent1, _, _, _, _ = run_model('vae_nb',
dataset1,
0,
0,
filename=plotname,
rep='vae1_nb')
latent2, _, _, _, _ = run_model('vae_nb',
dataset2,
0,
0,
filename=plotname,
rep='vae2_nb')
else:
latent1, _, _, _, _ = run_model('vae',
dataset1,
0,
0,
filename=plotname,
rep='vae1')
latent2, _, _, _, _ = run_model('vae',
dataset2,
0,
0,
filename=plotname,
rep='vae2')
for model_type in [
'vae', 'scanvi1', 'scanvi2', 'vae_nb', 'scanvi1_nb',
'scanvi2_nb'
]:
print(model_type)
latent, batch_indices, labels, keys, stats = run_model(
model_type,
gene_dataset,
dataset1,
dataset2,
filename=plotname,
rep='0')
res_jaccard = [
KNNJaccardIndex(latent1, latent2, latent, batch_indices, k)[0]
for k in KNeighbors
]
res_jaccard_score = np.sum(res_jaccard * K_int)
res_knn, res_knn_partial, res_kmeans, res_kmeans_partial = \
eval_latent(batch_indices=batch_indices, labels=labels, latent=latent, keys=keys,
labelled_idx=labelled_idx, unlabelled_idx=unlabelled_idx,
plotname=plotname + '.' + model_type, plotting=False, partial_only=False)
_, res_knn_partial1, _, res_kmeans_partial1 = \
eval_latent(batch_indices=batch_indices, labels=labels, latent=latent, keys=keys,
labelled_idx=(batch_indices == 0), unlabelled_idx=(batch_indices == 1),
plotname=plotname + '.' + model_type, plotting=False)
_, res_knn_partial2, _, res_kmeans_partial2 = \
eval_latent(batch_indices=batch_indices, labels=labels, latent=latent, keys=keys,
labelled_idx=(batch_indices == 1), unlabelled_idx=(batch_indices == 0),
plotname=plotname + '.' + model_type, plotting=False)
res = [res_knn[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_knn_partial[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_knn_partial1[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_knn_partial2[x] for x in ['asw', 'nmi', 'ari', 'ca', 'weighted ca']] + \
[res_kmeans[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
[res_kmeans_partial[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
[res_kmeans_partial1[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
[res_kmeans_partial2[x] for x in ['asw', 'nmi', 'ari', 'uca', 'weighted uca']] + \
res_jaccard + \
[res_jaccard_score, stats[0], stats[1], stats[2]]
f.write(model_type + (" %.4f" * 61 + "\n") % tuple(res))
g.write("%s\t" % (model_type) + 'all\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + 'p\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn_partial['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + 'p1\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn_partial1['clusteracc']) + "\n"))
g.write("%s\t" % (model_type) + 'p2\t' +
("%.4f\t" * len(gene_dataset.cell_types) %
tuple(res_knn_partial2['clusteracc']) + "\n"))
# for i in [1, 2, 3]:
# latent, batch_indices, labels, keys, stats = run_model(model_type, gene_dataset, dataset1, dataset2,
# filename=plotname, rep=str(i))
# res_jaccard, res_jaccard_score = KNNJaccardIndex(latent1, latent2, latent, batch_indices)
#
# res_knn, res_knn_partial, res_kmeans, res_kmeans_partial = \
# eval_latent(batch_indices=batch_indices, labels=labels, latent=latent, keys=keys,
# labelled_idx=labelled_idx, unlabelled_idx=unlabelled_idx,
# plotname=plotname + '.' + model_type, plotting=False,partial_only=False)
#
# _, res_knn_partial1, _, res_kmeans_partial1 = \
# eval_latent(batch_indices=batch_indices, labels=labels, latent=latent, keys=keys,
# labelled_idx=(batch_indices == 0), unlabelled_idx=(batch_indices == 1),
# plotname=plotname + '.' + model_type, plotting=False)
#
# _, res_knn_partial2, _, res_kmeans_partial2 = \
# eval_latent(batch_indices=batch_indices, labels=labels, latent=latent, keys=keys,
# labelled_idx=(batch_indices == 1), unlabelled_idx=(batch_indices == 0),
# plotname=plotname + '.' + model_type, plotting=False)
#
# res = [res_knn[x] for x in res_knn] + \
# [res_knn_partial[x] for x in res_knn_partial] + \
# [res_knn_partial1[x] for x in res_knn_partial1] + \
# [res_knn_partial2[x] for x in res_knn_partial2] + \
# [res_kmeans[x] for x in res_kmeans] + \
# [res_kmeans_partial[x] for x in res_kmeans_partial] + \
# [res_kmeans_partial1[x] for x in res_kmeans_partial1] + \
# [res_kmeans_partial2[x] for x in res_kmeans_partial2] + \
# res_jaccard + \
# [res_jaccard_score,stats[0], stats[1], stats[2]]
# f.write(model_type + (" %.4f" * 61 + "\n") % tuple(res))
elif models == 'writedata':
_, _, _, _, _ = run_model('writedata',
gene_dataset,
dataset1,
dataset2,
filename=plotname)
f.close()
g.close()