forked from stephenliu0423/PyDTI
-
Notifications
You must be signed in to change notification settings - Fork 0
/
nrlmf.py
179 lines (166 loc) · 7.86 KB
/
nrlmf.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
'''
[1] Yong Liu, Min Wu, Chunyan Miao, Peilin Zhao, Xiao-Li Li, "Neighborhood Regularized Logistic Matrix Factorization for Drug-target Interaction Prediction", under review.
'''
import numpy as np
from sklearn.metrics import precision_recall_curve, roc_curve
from sklearn.metrics import auc
class NRLMF:
def __init__(self, cfix=5, K1=5, K2=5, num_factors=10, theta=1.0, lambda_d=0.625, lambda_t=0.625, alpha=0.1, beta=0.1, max_iter=100):
self.cfix = int(cfix) # importance level for positive observations
self.K1 = int(K1)
self.K2 = int(K2)
self.num_factors = int(num_factors)
self.theta = float(theta)
self.lambda_d = float(lambda_d)
self.lambda_t = float(lambda_t)
self.alpha = float(alpha)
self.beta = float(beta)
self.max_iter = int(max_iter)
def AGD_optimization(self, seed=None):
if seed is None:
self.U = np.sqrt(1/float(self.num_factors))*np.random.normal(size=(self.num_drugs, self.num_factors))
self.V = np.sqrt(1/float(self.num_factors))*np.random.normal(size=(self.num_targets, self.num_factors))
else:
prng = np.random.RandomState(seed)
self.U = np.sqrt(1/float(self.num_factors))*prng.normal(size=(self.num_drugs, self.num_factors))
self.V = np.sqrt(1/float(self.num_factors))*prng.normal(size=(self.num_targets, self.num_factors))
dg_sum = np.zeros((self.num_drugs, self.U.shape[1]))
tg_sum = np.zeros((self.num_targets, self.V.shape[1]))
last_log = self.log_likelihood()
for t in range(self.max_iter):
dg = self.deriv(True)
dg_sum += np.square(dg)
vec_step_size = self.theta / np.sqrt(dg_sum)
self.U += vec_step_size * dg
tg = self.deriv(False)
tg_sum += np.square(tg)
vec_step_size = self.theta / np.sqrt(tg_sum)
self.V += vec_step_size * tg
curr_log = self.log_likelihood()
delta_log = (curr_log-last_log)/abs(last_log)
if abs(delta_log) < 1e-5:
break
last_log = curr_log
def deriv(self, drug):
if drug:
vec_deriv = np.dot(self.intMat, self.V)
else:
vec_deriv = np.dot(self.intMat.T, self.U)
A = np.dot(self.U, self.V.T)
A = np.exp(A)
A /= (A + self.ones)
A = self.intMat1 * A
if drug:
vec_deriv -= np.dot(A, self.V)
vec_deriv -= self.lambda_d*self.U+self.alpha*np.dot(self.DL, self.U)
else:
vec_deriv -= np.dot(A.T, self.U)
vec_deriv -= self.lambda_t*self.V+self.beta*np.dot(self.TL, self.V)
return vec_deriv
def log_likelihood(self):
loglik = 0
A = np.dot(self.U, self.V.T)
B = A * self.intMat
loglik += np.sum(B)
A = np.exp(A)
A += self.ones
A = np.log(A)
A = self.intMat1 * A
loglik -= np.sum(A)
loglik -= 0.5 * self.lambda_d * np.sum(np.square(self.U))+0.5 * self.lambda_t * np.sum(np.square(self.V))
loglik -= 0.5 * self.alpha * np.sum(np.diag((np.dot(self.U.T, self.DL)).dot(self.U)))
loglik -= 0.5 * self.beta * np.sum(np.diag((np.dot(self.V.T, self.TL)).dot(self.V)))
return loglik
def construct_neighborhood(self, drugMat, targetMat):
self.dsMat = drugMat - np.diag(np.diag(drugMat))
self.tsMat = targetMat - np.diag(np.diag(targetMat))
if self.K1 > 0:
S1 = self.get_nearest_neighbors(self.dsMat, self.K1)
self.DL = self.laplacian_matrix(S1)
S2 = self.get_nearest_neighbors(self.tsMat, self.K1)
self.TL = self.laplacian_matrix(S2)
else:
self.DL = self.laplacian_matrix(self.dsMat)
self.TL = self.laplacian_matrix(self.tsMat)
def laplacian_matrix(self, S):
x = np.sum(S, axis=0)
y = np.sum(S, axis=1)
L = 0.5*(np.diag(x+y) - (S+S.T)) # neighborhood regularization matrix
return L
def get_nearest_neighbors(self, S, size=5):
m, n = S.shape
X = np.zeros((m, n))
for i in xrange(m):
ii = np.argsort(S[i, :])[::-1][:min(size, n)]
X[i, ii] = S[i, ii]
return X
def fix_model(self, W, intMat, drugMat, targetMat, seed=None):
self.num_drugs, self.num_targets = intMat.shape
self.ones = np.ones((self.num_drugs, self.num_targets))
self.intMat = self.cfix*intMat*W
self.intMat1 = (self.cfix-1)*intMat*W + self.ones
x, y = np.where(self.intMat > 0)
self.train_drugs, self.train_targets = set(x.tolist()), set(y.tolist())
self.construct_neighborhood(drugMat, targetMat)
self.AGD_optimization(seed)
def predict_scores(self, test_data, N):
dinx = np.array(list(self.train_drugs))
DS = self.dsMat[:, dinx]
tinx = np.array(list(self.train_targets))
TS = self.tsMat[:, tinx]
scores = []
for d, t in test_data:
if d in self.train_drugs:
if t in self.train_targets:
val = np.sum(self.U[d, :]*self.V[t, :])
else:
jj = np.argsort(TS[t, :])[::-1][:N]
val = np.sum(self.U[d, :]*np.dot(TS[t, jj], self.V[tinx[jj], :]))/np.sum(TS[t, jj])
else:
if t in self.train_targets:
ii = np.argsort(DS[d, :])[::-1][:N]
val = np.sum(np.dot(DS[d, ii], self.U[dinx[ii], :])*self.V[t, :])/np.sum(DS[d, ii])
else:
ii = np.argsort(DS[d, :])[::-1][:N]
jj = np.argsort(TS[t, :])[::-1][:N]
v1 = DS[d, ii].dot(self.U[dinx[ii], :])/np.sum(DS[d, ii])
v2 = TS[t, jj].dot(self.V[tinx[jj], :])/np.sum(TS[t, jj])
val = np.sum(v1*v2)
scores.append(np.exp(val)/(1+np.exp(val)))
return np.array(scores)
def evaluation(self, test_data, test_label):
dinx = np.array(list(self.train_drugs))
DS = self.dsMat[:, dinx]
tinx = np.array(list(self.train_targets))
TS = self.tsMat[:, tinx]
scores = []
if self.K2 > 0:
for d, t in test_data:
if d in self.train_drugs:
if t in self.train_targets:
val = np.sum(self.U[d, :]*self.V[t, :])
else:
jj = np.argsort(TS[t, :])[::-1][:self.K2]
val = np.sum(self.U[d, :]*np.dot(TS[t, jj], self.V[tinx[jj], :]))/np.sum(TS[t, jj])
else:
if t in self.train_targets:
ii = np.argsort(DS[d, :])[::-1][:self.K2]
val = np.sum(np.dot(DS[d, ii], self.U[dinx[ii], :])*self.V[t, :])/np.sum(DS[d, ii])
else:
ii = np.argsort(DS[d, :])[::-1][:self.K2]
jj = np.argsort(TS[t, :])[::-1][:self.K2]
v1 = DS[d, ii].dot(self.U[dinx[ii], :])/np.sum(DS[d, ii])
v2 = TS[t, jj].dot(self.V[tinx[jj], :])/np.sum(TS[t, jj])
val = np.sum(v1*v2)
scores.append(np.exp(val)/(1+np.exp(val)))
elif self.K2 == 0:
for d, t in test_data:
val = np.sum(self.U[d, :]*self.V[t, :])
scores.append(np.exp(val)/(1+np.exp(val)))
prec, rec, thr = precision_recall_curve(test_label, np.array(scores))
aupr_val = auc(rec, prec)
fpr, tpr, thr = roc_curve(test_label, np.array(scores))
auc_val = auc(fpr, tpr)
return aupr_val, auc_val
def __str__(self):
return "Model: NRLMF, c:%s, K1:%s, K2:%s, r:%s, lambda_d:%s, lambda_t:%s, alpha:%s, beta:%s, theta:%s, max_iter:%s" % (self.cfix, self.K1, self.K2, self.num_factors, self.lambda_d, self.lambda_t, self.alpha, self.beta, self.theta, self.max_iter)