-
Notifications
You must be signed in to change notification settings - Fork 34
/
server-cao.py
124 lines (92 loc) · 4.33 KB
/
server-cao.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
import numpy as np
import pandas as pd
import scanpy.api as sc
sc.settings.verbosity = 3
import pickle
import sys
sys.path.append('/gpfs01/berens/user/dkobak/FIt-SNE')
from fast_tsne import fast_tsne
# LOAD AND PREPROCESS THE DATA
# data from https://oncoscape.v3.sttrcancer.org/atlas.gs.washington.edu.mouse.rna/downloads
from scipy.io import mmread
from scipy import sparse
import anndata
%time counts = mmread('big-data/cao/gene_count.txt').transpose() # 38 min!!
counts = sparse.csr_matrix(counts)
adata = anndata.AnnData(counts)
sc.pp.recipe_zheng17(adata, n_top_genes=2000) # 2 min
X = np.copy(adata.X)
X = X - X.mean(axis=0)
# Note: svd took extra ~60Gb RAM. scipy.sparse.svds would probably be better here.
# I did not want to use randomized algorithms for this paper.
%time U, s, V = np.linalg.svd(X, full_matrices=False) # 1 min 46s
U[:, np.sum(V,axis=1)<0] *= -1
X = np.dot(U, np.diag(s))
X = X[:, np.argsort(s)[::-1]][:,:50]
pickle.dump(X, open('big-pickles/cao-pca.pickle', 'wb'))
# read in cluster labels
meta = pd.read_csv('big-data/cao/cell_annotate.csv')
clusters = meta['Main_Cluster'].values.copy()
clusters[~np.isnan(clusters)] -= 1
clusters[np.isnan(clusters)] = -1
clusters = clusters.astype(int)
# cluster names
clusterNames = meta['Main_cell_type'][np.unique(clusters, return_index=True)[1][1:]].values.copy()
# t-SNE embedding from the paper
Zpaper = np.concatenate((meta['Main_cluster_tsne_1'].values[:,np.newaxis],
meta['Main_cluster_tsne_2'].values[:,np.newaxis]), axis=1)
# DOWNSAMPLE AND RUN t-SNE
X = pickle.load(open('big-pickles/cao-pca.pickle', 'rb'))
PCAinit = X[:,:2] / np.std(X[:,0]) * 0.0001
np.random.seed(42)
ind25k = np.random.choice(X.shape[0], 25000, replace=False)
Z25k = fast_tsne(X[ind25k,:], perplexity_list=[30,int(25000/100)],
initialization = PCAinit[ind25k,:], seed=42,
learning_rate = 25000/12)
def downsampled_nn(X, Z, downsampled_ind, batchsize=1000, knn=10):
ind_rest = np.where(~np.isin(np.arange(X.shape[0]), downsampled_ind))[0]
steps = int(np.ceil(ind_rest.size/batchsize))
positions = np.zeros((X.shape[0], 2))
positions[downsampled_ind,:] = Z
def pdist2(A,B):
return np.sum(A**2,axis=1)[:, None] + np.sum(B**2, axis=1)[None, :] - 2 * A @ B.T
for i in range(steps):
print('.', end='', flush=True)
if (i+1)%100==0:
print('', flush=True)
endind = np.min(((i+1)*batchsize, ind_rest.size))
batch = ind_rest[i*batchsize:endind]
D = pdist2(X[batch, :], X[downsampled_ind,:])
ind = np.argpartition(D, knn)[:, :knn]
for i in range(batch.size):
positions[batch[i],:] = np.median(Z[ind[i,:],:], axis=0)
print('', flush=True)
return positions
%time positions = downsampled_nn(X, Z25k, ind25k, batchsize=10000) # 14 min
pickle.dump([Z25k, positions], open("big-pickles/cao-downsampling.pickle", "wb"))
# RUN T-SNE VARIANTS ON THE FULL DATA SET
X = pickle.load(open('big-pickles/cao-pca.pickle', 'rb'))
Z25k, positions = pickle.load(open("big-pickles/cao-downsampling.pickle", "rb"))
Zs = {}
init25k = positions/np.std(positions[:,0]) * 0.0001
%time Z = fast_tsne(X, perplexity=30, initialization=init25k, late_exag_coeff=4, start_late_exag_iter=250, learning_rate=X.shape[0]/12, seed=42, load_affinities='save') # 26min
Zs['mine'] = Z
Z = fast_tsne(X, perplexity=30, initialization=init25k, learning_rate=X.shape[0]/12, seed=42, load_affinities='load')
Zs['noexagg'] = Z
%time Z = fast_tsne(X, perplexity=30, initialization=PCAinit, late_exag_coeff=4, start_late_exag_iter=250, learning_rate=X.shape[0]/12, seed=42, load_affinities='load')
Zs['pcainit'] = Z
Z = fast_tsne(X, perplexity=30, initialization=PCAinit, learning_rate=X.shape[0]/12, seed=42, load_affinities='load')
Zs['noexagg-pcainit'] = Z
Z = fast_tsne(X, perplexity=30, late_exag_coeff=4, start_late_exag_iter=250, learning_rate=X.shape[0]/12, seed=42, load_affinities='load')
Zs['randinit'] = Z
Z = fast_tsne(X, perplexity=30, learning_rate=1000, seed=42, load_affinities='load')
Zs['scanpy'] = Z
Z = fast_tsne(X, perplexity=30, learning_rate=X.shape[0]/12, seed=42, load_affinities='load')
Zs['belkina'] = Z
Z = fast_tsne(X, perplexity=30, seed=42, load_affinities='load')
Zs['default'] = Z
import umap
%time Z = umap.UMAP(random_state=1).fit_transform(X) # 2h 5min
Zs['umap'] = Z
Zs['paper'] = Zpaper
pickle.dump([Zs, clusters, clusterNames], open("big-pickles/cao-tsne.pickle", "wb"))