-
Notifications
You must be signed in to change notification settings - Fork 2
/
pyWNN.py
235 lines (193 loc) · 9.09 KB
/
pyWNN.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
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
import scanpy as sc
import numpy as np
from sklearn import preprocessing
from scipy.sparse import csr_matrix, lil_matrix, diags
import time
def get_nearestneighbor(knn, neighbor=1):
'''For each row of knn, returns the column with the lowest value
I.e. the nearest neighbor'''
indices = knn.indices
indptr = knn.indptr
data = knn.data
nn_idx = []
for i in range(knn.shape[0]):
cols = indices[indptr[i]:indptr[i+1]]
rowvals = data[indptr[i]:indptr[i+1]]
idx = np.argsort(rowvals)
nn_idx.append(cols[idx[neighbor-1]])
return(np.array(nn_idx))
def compute_bw(knn_adj, embedding, n_neighbors=20):
intersect = knn_adj.dot(knn_adj.T)
indices = intersect.indices
indptr = intersect.indptr
data = intersect.data
data = data / ((n_neighbors*2) - data)
bandwidth = []
for i in range(intersect.shape[0]):
cols = indices[indptr[i]:indptr[i+1]]
rowvals = data[indptr[i]:indptr[i+1]]
idx = np.argsort(rowvals)
valssort = rowvals[idx]
numinset = len(cols)
if numinset<n_neighbors:
sys.exit('Fewer than 20 cells with Jacard sim > 0')
else:
curval = valssort[n_neighbors]
for num in range(n_neighbors, numinset):
if valssort[num]!=curval:
break
else:
num+=1
minjacinset = cols[idx][:num]
if num <n_neighbors:
print('shouldnt end up here')
sys.exit(-1)
else:
euc_dist = ((embedding[minjacinset,:]-embedding[i,:])**2).sum(axis=1)**.5
euc_dist_sorted = np.sort(euc_dist)[::-1]
bandwidth.append( np.mean(euc_dist_sorted[:n_neighbors]) )
return(np.array(bandwidth))
def compute_affinity(dist_to_predict, dist_to_nn, bw):
affinity = dist_to_predict-dist_to_nn
affinity[affinity<0]=0
affinity = affinity * -1
affinity = np.exp(affinity / (bw-dist_to_nn))
return(affinity)
def dist_from_adj(adjacency, embed1, embed2, nndist1, nndist2):
dist1 = lil_matrix(adjacency.shape)
dist2 = lil_matrix(adjacency.shape)
count = 0
indices = adjacency.indices
indptr = adjacency.indptr
ncells = adjacency.shape[0]
tic = time.perf_counter()
for i in range(ncells):
for j in range(indptr[i], indptr[i+1]):
col = indices[j]
a = (((embed1[i,:] - embed1[col,:])**2).sum()**.5) - nndist1[i]
if a == 0: dist1[i,col] = np.nan
else: dist1[i,col] = a
b = (((embed2[i,:] - embed2[col,:])**2).sum()**.5) - nndist2[i]
if b == 0: dist2[i,col] = np.nan
else: dist2[i,col] = b
if (i % 2000) == 0:
toc = time.perf_counter()
print('%d out of %d %.2f seconds elapsed' % (i, ncells, toc-tic))
return(csr_matrix(dist1), csr_matrix(dist2))
def select_topK(dist, n_neighbors=20):
indices = dist.indices
indptr = dist.indptr
data = dist.data
nrows = dist.shape[0]
final_data = []
final_col_ind = []
tic = time.perf_counter()
for i in range(nrows):
cols = indices[indptr[i]:indptr[i+1]]
rowvals = data[indptr[i]:indptr[i+1]]
idx = np.argsort(rowvals)
final_data.append(rowvals[idx[(-1*n_neighbors):]])
final_col_ind.append(cols[idx[(-1*n_neighbors):]])
final_data = np.concatenate(final_data)
final_col_ind = np.concatenate(final_col_ind)
final_row_ind = np.tile(np.arange(nrows), (n_neighbors, 1)).reshape(-1, order='F')
result = csr_matrix((final_data, (final_row_ind, final_col_ind)), shape=(nrows, dist.shape[1]))
return(result)
class pyWNN():
def __init__(self, adata, reps=['X_pca', 'X_apca'], n_neighbors=20, npcs=[20, 20], seed=14, distances=None):
"""\
Class for running weighted nearest neighbors analysis as described in Hao
et al 2021.
"""
self.seed = seed
np.random.seed(seed)
if len(reps)>2:
sys.exit('WNN currently only implemented for 2 modalities')
self.adata = adata.copy()
self.reps = [r+'_norm' for r in reps]
self.npcs = npcs
for (i,r) in enumerate(reps):
self.adata.obsm[self.reps[i]] = preprocessing.normalize(adata.obsm[r][:,0:npcs[i]])
self.n_neighbors = n_neighbors
if distances is None:
print('Computing KNN distance matrices using default Scanpy implementation')
sc.pp.neighbors(self.adata, n_neighbors=n_neighbors, n_pcs=npcs[0], use_rep=self.reps[0], metric='euclidean', key_added='1')
sc.pp.neighbors(self.adata, n_neighbors=n_neighbors, n_pcs=npcs[1], use_rep=self.reps[1], metric='euclidean', key_added='2')
sc.pp.neighbors(self.adata, n_neighbors=200, n_pcs=npcs[0], use_rep=self.reps[0], metric='euclidean', key_added='1_200')
sc.pp.neighbors(self.adata, n_neighbors=200, n_pcs=npcs[1], use_rep=self.reps[1], metric='euclidean', key_added='2_200')
self.distances = ['1_distances', '2_distances', '1_200_distances', '2_200_distances']
else:
self.distances = distances
for d in self.distances:
if type(self.adata.obsp[d]) is not csr_matrix:
self.adata.obsp[d] = csr_matrix(self.adata.obsp[d])
self.NNdist = []
self.NNidx = []
self.NNadjacency = []
self.BWs = []
for (i,r) in enumerate(self.reps):
nn = get_nearestneighbor(self.adata.obsp[self.distances[i]])
dist_to_nn = ((self.adata.obsm[r]-self.adata.obsm[r][nn, :])**2).sum(axis=1)**.5
nn_adj = (self.adata.obsp[self.distances[i]]>0).astype(int)
nn_adj_wdiag = nn_adj.copy()
nn_adj_wdiag.setdiag(1)
bw = compute_bw(nn_adj_wdiag, self.adata.obsm[r], n_neighbors=self.n_neighbors)
self.NNidx.append(nn)
self.NNdist.append(dist_to_nn)
self.NNadjacency.append(nn_adj)
self.BWs.append(bw)
self.weights = []
self.WNN = None
def compute_weights(self):
cmap = {0:1, 1:0}
affinity_ratios = []
self.within = []
self.cross = []
for (i,r) in enumerate(self.reps):
within_predict = self.NNadjacency[i].dot(self.adata.obsm[r]) / (self.n_neighbors-1)
cross_predict = self.NNadjacency[cmap[i]].dot(self.adata.obsm[r]) / (self.n_neighbors-1)
within_predict_dist = ((self.adata.obsm[r] - within_predict)**2).sum(axis=1)**.5
cross_predict_dist = ((self.adata.obsm[r] - cross_predict)**2).sum(axis=1)**.5
within_affinity = compute_affinity(within_predict_dist, self.NNdist[i], self.BWs[i])
cross_affinity = compute_affinity(cross_predict_dist, self.NNdist[i], self.BWs[i])
affinity_ratios.append(within_affinity / (cross_affinity + 0.0001))
self.within.append(within_predict_dist)
self.cross.append(cross_predict_dist)
self.weights.append( 1 / (1+ np.exp(affinity_ratios[1]-affinity_ratios[0])) )
self.weights.append( 1 - self.weights[0] )
def compute_wnn(self, adata):
print('Computing modality weights')
self.compute_weights()
union_adj_mat = ((self.adata.obsp[self.distances[2]]+self.adata.obsp[self.distances[3]]) > 0).astype(int)
print('Computing weighted distances for union of 200 nearest neighbors between modalities')
full_dists = dist_from_adj(union_adj_mat, self.adata.obsm[self.reps[0]], self.adata.obsm[self.reps[1]],
self.NNdist[0], self.NNdist[1])
weighted_dist = csr_matrix(union_adj_mat.shape)
for (i,dist) in enumerate(full_dists):
dist = diags(-1 / (self.BWs[i] - self.NNdist[i]), format='csr').dot(dist)
dist.data = np.exp(dist.data)
ind = np.isnan(dist.data)
dist.data[ind] = 1
dist = diags(self.weights[i]).dot(dist)
weighted_dist += dist
print('Selecting top K neighbors')
self.WNN = select_topK(weighted_dist, n_neighbors=self.n_neighbors)
WNNdist = self.WNN.copy()
x = (1-WNNdist.data) / 2
x[x<0]=0
x[x>1]=1
WNNdist.data = np.sqrt(x)
self.WNNdist = WNNdist
adata.obsp['WNN'] = self.WNN
adata.obsp['WNN_distance'] = self.WNNdist
adata.obsm[self.reps[0]] = self.adata.obsm[self.reps[0]]
adata.obsm[self.reps[1]] = self.adata.obsm[self.reps[1]]
adata.uns['WNN'] = {'connectivities_key': 'WNN',
'distances_key': 'WNN_distance',
'params': {'n_neighbors': self.n_neighbors,
'method': 'WNN',
'random_state': self.seed,
'metric': 'euclidean',
'use_rep': self.reps[0],
'n_pcs': self.npcs[0]}}
return(adata)