|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: UTF-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
""" |
|
5
|
|
|
Compute semantic similarities between GO terms. Borrowed from book chapter from |
|
6
|
|
|
Christophe Dessimoz (thanks). For details, please check out: |
|
7
|
|
|
notebooks/semantic_similarity.ipynb |
|
8
|
|
|
""" |
|
9
|
|
|
|
|
10
|
|
|
import math |
|
11
|
|
|
from collections import Counter |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class TermCounts: |
|
15
|
|
|
''' |
|
16
|
|
|
TermCounts counts the term counts for each |
|
17
|
|
|
''' |
|
18
|
|
|
def __init__(self, go, annots): |
|
19
|
|
|
''' |
|
20
|
|
|
Initialise the counts and |
|
21
|
|
|
''' |
|
22
|
|
|
# Backup |
|
23
|
|
|
self._go = go |
|
24
|
|
|
|
|
25
|
|
|
# Initialise the counters |
|
26
|
|
|
self._counts = Counter() |
|
27
|
|
|
self._aspect_counts = Counter() |
|
28
|
|
|
|
|
29
|
|
|
# Fill the counters... |
|
30
|
|
|
self._count_terms(go, annots) |
|
31
|
|
|
|
|
32
|
|
|
def _count_terms(self, go, annots): |
|
33
|
|
|
''' |
|
34
|
|
|
Fills in the counts and overall aspect counts. |
|
35
|
|
|
''' |
|
36
|
|
|
for gene, terms in annots.items(): |
|
37
|
|
|
# Make a union of all the terms for a gene, if term parents are |
|
38
|
|
|
# propagated but they won't get double-counted for the gene |
|
39
|
|
|
allterms = set(terms) |
|
40
|
|
|
for go_id in terms: |
|
41
|
|
|
allterms |= go[go_id].get_all_parents() |
|
42
|
|
|
for p in allterms: |
|
43
|
|
|
self._counts[p] += 1 |
|
44
|
|
|
|
|
45
|
|
|
for go_id, c in self._counts.items(): |
|
46
|
|
|
# Group by namespace |
|
47
|
|
|
namespace = go[go_id].namespace |
|
48
|
|
|
self._aspect_counts[namespace] += c |
|
49
|
|
|
|
|
50
|
|
|
def get_count(self, go_id): |
|
51
|
|
|
''' |
|
52
|
|
|
Returns the count of that GO term observed in the annotations. |
|
53
|
|
|
''' |
|
54
|
|
|
return self._counts[go_id] |
|
55
|
|
|
|
|
56
|
|
|
def get_total_count(self, aspect): |
|
57
|
|
|
''' |
|
58
|
|
|
Gets the total count that's been precomputed. |
|
59
|
|
|
''' |
|
60
|
|
|
return self._aspect_counts[aspect] |
|
61
|
|
|
|
|
62
|
|
|
def get_term_freq(self, go_id): |
|
63
|
|
|
''' |
|
64
|
|
|
Returns the frequency at which a particular GO term has |
|
65
|
|
|
been observed in the annotations. |
|
66
|
|
|
''' |
|
67
|
|
|
try: |
|
68
|
|
|
namespace = self._go[go_id].namespace |
|
69
|
|
|
freq = float(self.get_count(go_id)) / float(self.get_total_count(namespace)) |
|
70
|
|
|
#print self.get_count(go_id), self.get_total_count(namespace), freq |
|
71
|
|
|
except ZeroDivisionError: |
|
72
|
|
|
freq = 0 |
|
73
|
|
|
|
|
74
|
|
|
return freq |
|
75
|
|
|
|
|
76
|
|
|
|
|
77
|
|
|
def ic(go_id, termcounts): |
|
78
|
|
|
''' |
|
79
|
|
|
Calculates the information content of a GO term. |
|
80
|
|
|
''' |
|
81
|
|
|
# Get the observed frequency of the GO term |
|
82
|
|
|
freq = termcounts.get_term_freq(go_id) |
|
83
|
|
|
|
|
84
|
|
|
# Calculate the information content (i.e., -log("freq of GO term") |
|
85
|
|
|
return -1.0 * math.log(freq) if freq else 0 |
|
86
|
|
|
|
|
87
|
|
|
|
|
88
|
|
|
def resnik_sim(go_id1, go_id2, go, termcounts): |
|
89
|
|
|
''' |
|
90
|
|
|
Computes Resnik's similarity measure. |
|
91
|
|
|
''' |
|
92
|
|
|
msca = deepest_common_ancestor([go_id1, go_id2], go) |
|
93
|
|
|
return ic(msca, termcounts) |
|
94
|
|
|
|
|
95
|
|
|
|
|
96
|
|
|
def lin_sim(go_id1, go_id2, go, termcounts): |
|
97
|
|
|
''' |
|
98
|
|
|
Computes Lin's similarity measure. |
|
99
|
|
|
''' |
|
100
|
|
|
sim_r = resnik_sim(go_id1, go_id2, go, termcounts) |
|
101
|
|
|
|
|
102
|
|
|
return (-2*sim_r)/(ic(go_id1, termcounts) + ic(go_id2, termcounts)) |
|
103
|
|
|
|
|
104
|
|
|
|
|
105
|
|
|
def common_parent_go_ids(terms, go): |
|
106
|
|
|
''' |
|
107
|
|
|
This function finds the common ancestors in the GO |
|
108
|
|
|
tree of the list of terms in the input. |
|
109
|
|
|
''' |
|
110
|
|
|
# Find candidates from first |
|
111
|
|
|
rec = go[terms[0]] |
|
112
|
|
|
candidates = rec.get_all_parents() |
|
113
|
|
|
candidates.update({terms[0]}) |
|
114
|
|
|
|
|
115
|
|
|
# Find intersection with second to nth term |
|
116
|
|
|
for term in terms[1:]: |
|
117
|
|
|
rec = go[term] |
|
118
|
|
|
parents = rec.get_all_parents() |
|
119
|
|
|
parents.update({term}) |
|
120
|
|
|
|
|
121
|
|
|
# Find the intersection with the candidates, and update. |
|
122
|
|
|
candidates.intersection_update(parents) |
|
123
|
|
|
|
|
124
|
|
|
return candidates |
|
125
|
|
|
|
|
126
|
|
|
|
|
127
|
|
|
def deepest_common_ancestor(terms, go): |
|
128
|
|
|
''' |
|
129
|
|
|
This function gets the nearest common ancestor |
|
130
|
|
|
using the above function. |
|
131
|
|
|
Only returns single most specific - assumes unique exists. |
|
132
|
|
|
''' |
|
133
|
|
|
# Take the element at maximum depth. |
|
134
|
|
|
return max(common_parent_go_ids(terms, go), key=lambda t: go[t].depth) |
|
135
|
|
|
|
|
136
|
|
|
|
|
137
|
|
|
def min_branch_length(go_id1, go_id2, go): |
|
138
|
|
|
''' |
|
139
|
|
|
Finds the minimum branch length between two terms in the GO DAG. |
|
140
|
|
|
''' |
|
141
|
|
|
# First get the deepest common ancestor |
|
142
|
|
|
dca = deepest_common_ancestor([go_id1, go_id2], go) |
|
143
|
|
|
|
|
144
|
|
|
# Then get the distance from the DCA to each term |
|
145
|
|
|
dca_depth = go[dca].depth |
|
146
|
|
|
d1 = go[go_id1].depth - dca_depth |
|
147
|
|
|
d2 = go[go_id2].depth - dca_depth |
|
148
|
|
|
|
|
149
|
|
|
# Return the total distance - i.e., to the deepest common ancestor and back. |
|
150
|
|
|
return d1 + d2 |
|
151
|
|
|
|
|
152
|
|
|
|
|
153
|
|
|
def semantic_distance(go_id1, go_id2, go): |
|
154
|
|
|
''' |
|
155
|
|
|
Finds the semantic distance (minimum number of connecting branches) |
|
156
|
|
|
between two GO terms. |
|
157
|
|
|
''' |
|
158
|
|
|
return min_branch_length(go_id1, go_id2, go) |
|
159
|
|
|
|
|
160
|
|
|
|
|
161
|
|
|
def semantic_similarity(go_id1, go_id2, go): |
|
162
|
|
|
''' |
|
163
|
|
|
Finds the semantic similarity (inverse of the semantic distance) |
|
164
|
|
|
between two GO terms. |
|
165
|
|
|
''' |
|
166
|
|
|
return 1.0 / float(semantic_distance(go_id1, go_id2, go)) |
|
167
|
|
|
|