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
|
"""
==========================================================================
Gaussian processes on discrete data structures
==========================================================================
This example illustrates the use of Gaussian processes for regression and
classification tasks on data that are not in fixed-length feature vector form.
This is achieved through the use of kernel functions that operates directly
on discrete structures such as variable-length sequences, trees, and graphs.
Specifically, here the input variables are some gene sequences stored as
variable-length strings consisting of letters 'A', 'T', 'C', and 'G',
while the output variables are floating point numbers and True/False labels
in the regression and classification tasks, respectively.
A kernel between the gene sequences is defined using R-convolution [1]_ by
integrating a binary letter-wise kernel over all pairs of letters among a pair
of strings.
This example will generate three figures.
In the first figure, we visualize the value of the kernel, i.e. the similarity
of the sequences, using a colormap. Brighter color here indicates higher
similarity.
In the second figure, we show some regression result on a dataset of 6
sequences. Here we use the 1st, 2nd, 4th, and 5th sequences as the training set
to make predictions on the 3rd and 6th sequences.
In the third figure, we demonstrate a classification model by training on 6
sequences and make predictions on another 5 sequences. The ground truth here is
simply whether there is at least one 'A' in the sequence. Here the model makes
four correct classifications and fails on one.
.. [1] Haussler, D. (1999). Convolution kernels on discrete structures
(Vol. 646). Technical report, Department of Computer Science, University
of California at Santa Cruz.
"""
# %%
import numpy as np
from sklearn.base import clone
from sklearn.gaussian_process import GaussianProcessClassifier, GaussianProcessRegressor
from sklearn.gaussian_process.kernels import GenericKernelMixin, Hyperparameter, Kernel
class SequenceKernel(GenericKernelMixin, Kernel):
"""
A minimal (but valid) convolutional kernel for sequences of variable
lengths."""
def __init__(self, baseline_similarity=0.5, baseline_similarity_bounds=(1e-5, 1)):
self.baseline_similarity = baseline_similarity
self.baseline_similarity_bounds = baseline_similarity_bounds
@property
def hyperparameter_baseline_similarity(self):
return Hyperparameter(
"baseline_similarity", "numeric", self.baseline_similarity_bounds
)
def _f(self, s1, s2):
"""
kernel value between a pair of sequences
"""
return sum(
[1.0 if c1 == c2 else self.baseline_similarity for c1 in s1 for c2 in s2]
)
def _g(self, s1, s2):
"""
kernel derivative between a pair of sequences
"""
return sum([0.0 if c1 == c2 else 1.0 for c1 in s1 for c2 in s2])
def __call__(self, X, Y=None, eval_gradient=False):
if Y is None:
Y = X
if eval_gradient:
return (
np.array([[self._f(x, y) for y in Y] for x in X]),
np.array([[[self._g(x, y)] for y in Y] for x in X]),
)
else:
return np.array([[self._f(x, y) for y in Y] for x in X])
def diag(self, X):
return np.array([self._f(x, x) for x in X])
def is_stationary(self):
return False
def clone_with_theta(self, theta):
cloned = clone(self)
cloned.theta = theta
return cloned
kernel = SequenceKernel()
# %%
# Sequence similarity matrix under the kernel
# ===========================================
import matplotlib.pyplot as plt
X = np.array(["AGCT", "AGC", "AACT", "TAA", "AAA", "GAACA"])
K = kernel(X)
D = kernel.diag(X)
plt.figure(figsize=(8, 5))
plt.imshow(np.diag(D**-0.5).dot(K).dot(np.diag(D**-0.5)))
plt.xticks(np.arange(len(X)), X)
plt.yticks(np.arange(len(X)), X)
plt.title("Sequence similarity under the kernel")
plt.show()
# %%
# Regression
# ==========
X = np.array(["AGCT", "AGC", "AACT", "TAA", "AAA", "GAACA"])
Y = np.array([1.0, 1.0, 2.0, 2.0, 3.0, 3.0])
training_idx = [0, 1, 3, 4]
gp = GaussianProcessRegressor(kernel=kernel)
gp.fit(X[training_idx], Y[training_idx])
plt.figure(figsize=(8, 5))
plt.bar(np.arange(len(X)), gp.predict(X), color="b", label="prediction")
plt.bar(training_idx, Y[training_idx], width=0.2, color="r", alpha=1, label="training")
plt.xticks(np.arange(len(X)), X)
plt.title("Regression on sequences")
plt.legend()
plt.show()
# %%
# Classification
# ==============
X_train = np.array(["AGCT", "CGA", "TAAC", "TCG", "CTTT", "TGCT"])
# whether there are 'A's in the sequence
Y_train = np.array([True, True, True, False, False, False])
gp = GaussianProcessClassifier(kernel)
gp.fit(X_train, Y_train)
X_test = ["AAA", "ATAG", "CTC", "CT", "C"]
Y_test = [True, True, False, False, False]
plt.figure(figsize=(8, 5))
plt.scatter(
np.arange(len(X_train)),
[1.0 if c else -1.0 for c in Y_train],
s=100,
marker="o",
edgecolor="none",
facecolor=(1, 0.75, 0),
label="training",
)
plt.scatter(
len(X_train) + np.arange(len(X_test)),
[1.0 if c else -1.0 for c in Y_test],
s=100,
marker="o",
edgecolor="none",
facecolor="r",
label="truth",
)
plt.scatter(
len(X_train) + np.arange(len(X_test)),
[1.0 if c else -1.0 for c in gp.predict(X_test)],
s=100,
marker="x",
facecolor="b",
linewidth=2,
label="prediction",
)
plt.xticks(np.arange(len(X_train) + len(X_test)), np.concatenate((X_train, X_test)))
plt.yticks([-1, 1], [False, True])
plt.title("Classification on sequences")
plt.legend()
plt.show()
|
