"""Simple association model
The core class here is AssociationSet, a holder for a set of
associations between entities such as genes and ontology
classes. AssociationSets can also be throught of as subsuming
traditional 'gene sets'
The model is deliberately simple, and does not seek to represent
metadata about the association - it is assumed that this is handled
upstream. See the assoc_factory module for details - this allows the
client to create an association set based on various criteria such as
taxa of interest or evidence criteria.
"""
import logging
import scipy.stats # TODO - move
import scipy as sp # TODO - move
import pandas as pd
logger = logging.getLogger(__name__)
class UnknownSubjectException(Exception):
pass
[docs]class AssociationSet():
"""An object that represents a collection of associations
NOTE: the intention is that this class can be subclassed to provide
either high-efficiency implementations, or implementations backed by services or external stores.
The default implementation is in-memory.
"""
def __init__(self, ontology=None, association_map=None, subject_label_map=None, meta=None):
"""
NOTE: in general you do not need to call this yourself. See assoc_factory
initializes an association set, which minimally consists of:
- an ontology (e.g. GO, HP)
- a map between subjects (e.g genes) and sets/lists of term IDs
"""
self.ontology = ontology
self.association_map = association_map
self.subject_label_map = subject_label_map
self.subject_to_inferred_map = {}
self.meta = meta # TODO
self.associations_by_subj = None
self.associations_by_subj_obj = None
self.strict = False
self.index()
if self.association_map is None:
self.association_map = {}
logger.info("Created {}".format(self))
def __str__(self):
imap = self.subject_to_inferred_map
return "AssocSet |S|={} |S->I|={}".format(len(imap.keys()), len(imap.items()))
[docs] def index(self):
"""
Creates indexes based on inferred terms.
You do not need to call this yourself; called on initialization
"""
self.subjects = list(self.association_map.keys())
# ensure annotations unique
for (subj,terms) in self.association_map.items():
self.association_map[subj] = list(set(self.association_map[subj]))
logger.info("Indexing {} items".format(len(self.subjects)))
n = 0
all_objs = set()
for (subj,terms) in self.association_map.items():
ancs = self.termset_ancestors(terms)
all_objs.update(ancs)
self.subject_to_inferred_map[subj] = ancs
n = n+1
if n<5:
logger.debug(" Indexed: {} -> {}".format(subj, ancs))
elif n == 6:
logger.info("[TRUNCATING>5]....")
self.objects = all_objs
[docs] def inferred_types(self, subj):
"""
Returns: set of reflexive inferred types for a subject.
E.g. if a gene is directly associated with terms A and B, and these terms have ancestors C, D and E
then the set returned will be {A,B,C,D,E}
Arguments
---------
subj - ID string
Returns: set of class IDs
"""
if subj in self.subject_to_inferred_map:
return self.subject_to_inferred_map[subj]
if self.strict:
raise UnknownSubjectException(subj)
else:
return set([])
[docs] def termset_ancestors(self, terms):
"""
reflexive ancestors
Arguments
---------
terms - a set or list of class IDs
Returns: set of class IDs
"""
ancs = set()
for term in terms:
ancs = ancs.union(self.ontology.ancestors(term))
return ancs.union(set(terms))
[docs] def query_associations(self, subjects=None, infer_subjects=True, include_xrefs=True):
"""
Query for a set of associations.
Note: only a minimal association model is stored, so all results are returned as (subject_id,class_id) tuples
Arguments:
subjects: list
list of subjects (e.g. genes, diseases) used to query associations. Any association to one of these subjects or
a descendant of these subjects (assuming infer_subjects=True) are returned.
infer_subjects: boolean (default true)
See above
include_xrefs: boolean (default true)
If true, then expand inferred subject set to include all xrefs of those subjects.
Example: if a high level disease node (e.g. DOID:14330 Parkinson disease) is specified, then the default behavior
(infer_subjects=True, include_xrefs=True) and the ontology includes DO, results will include associations from
both descendant DOID classes, and all xrefs (e.g. OMIM)
"""
if subjects is None:
subjects = []
mset = set()
if infer_subjects:
for subj in subjects:
mset.update(self.ontology.descendants(subj))
mset.update(set(subjects))
if include_xrefs:
xset = set()
for m in mset:
xrefs = self.ontology.xrefs(m, bidirectional=True)
if xrefs is not None:
xset.update(xrefs)
mset.update(xset)
logger.debug("Matching subjects: {}".format(mset))
mset = mset.intersection(self.subjects)
logger.debug("Matching subjects with anns: {}".format(mset))
amap = self.association_map
results = []
for m in mset:
if m in amap:
for t in amap[m]:
results.append( (m,t) )
return results
[docs] def annotations(self, subject_id):
"""
Returns a list of classes used to describe a subject
@Deprecated: use objects_for_subject
"""
if subject_id in self.association_map:
return self.association_map[subject_id]
else:
return []
[docs] def objects_for_subject(self, subject_id):
"""
Returns a list of classes used to describe a subject
"""
if subject_id in self.association_map:
return self.association_map[subject_id]
else:
return []
[docs] def query(self, terms=None, negated_terms=None):
"""
Basic boolean query, using inference.
Arguments:
- terms: list
list of class ids. Returns the set of subjects that have at least one inferred annotation to each of the specified classes.
- negated_terms: list
list of class ids. Filters the set of subjects so that there are no inferred annotations to any of the specified classes
"""
if terms is None:
terms = []
matches_all = 'owl:Thing' in terms
if negated_terms is None:
negated_terms = []
termset = set(terms)
negated_termset = set(negated_terms)
matches = []
n_terms = len(termset)
for subj in self.subjects:
if matches_all or len(termset.intersection(self.inferred_types(subj))) == n_terms:
if len(negated_termset.intersection(self.inferred_types(subj))) == 0:
matches.append(subj)
return matches
[docs] def query_intersections(self, x_terms=None, y_terms=None, symmetric=False):
"""
Query for intersections of terms in two lists
Return a list of intersection result objects with keys:
- x : term from x
- y : term from y
- c : count of intersection
- j : jaccard score
"""
if x_terms is None:
x_terms = []
if y_terms is None:
y_terms = []
xset = set(x_terms)
yset = set(y_terms)
zset = xset.union(yset)
# first built map of gene->termClosure.
# this could be calculated ahead of time for all g,
# but this may be space-expensive. TODO: benchmark
gmap={}
for z in zset:
gmap[z] = []
for subj in self.subjects:
ancs = self.inferred_types(subj)
for a in ancs.intersection(zset):
gmap[a].append(subj)
for z in zset:
gmap[z] = set(gmap[z])
ilist = []
for x in x_terms:
for y in y_terms:
if not symmetric or x<y:
shared = gmap[x].intersection(gmap[y])
union = gmap[x].union(gmap[y])
j = 0
if len(union)>0:
j = len(shared) / len(union)
ilist.append({'x':x,'y':y,'shared':shared, 'c':len(shared), 'j':j})
return ilist
[docs] @staticmethod
def intersectionlist_to_matrix(ilist, xterms, yterms):
"""
WILL BE DEPRECATED
Replace with method to return pandas dataframe
"""
z = [ [0] * len(xterms) for i1 in range(len(yterms)) ]
xmap = {}
xi = 0
for x in xterms:
xmap[x] = xi
xi = xi+1
ymap = {}
yi = 0
for y in yterms:
ymap[y] = yi
yi = yi+1
for i in ilist:
z[ymap[i['y']]][xmap[i['x']]] = i['j']
logger.debug("Z={}".format(z))
return (z,xterms,yterms)
[docs] def as_dataframe(self, fillna=True, subjects=None):
"""
Return association set as pandas DataFrame
Each row is a subject (e.g. gene)
Each column is the inferred class used to describe the subject
"""
entries = []
selected_subjects = self.subjects
if subjects is not None:
selected_subjects = subjects
for s in selected_subjects:
vmap = {}
for c in self.inferred_types(s):
vmap[c] = 1
entries.append(vmap)
logger.debug("Creating DataFrame")
df = pd.DataFrame(entries, index=selected_subjects)
if fillna:
logger.debug("Performing fillna...")
df = df.fillna(0)
return df
[docs] def label(self, id):
"""
return label for a subject id
Will make use of both the ontology and the association set
"""
if self.ontology is not None:
label = self.ontology.label(id)
if label is not None:
return label
if self.subject_label_map is not None and id in self.subject_label_map:
return self.subject_label_map[id]
return None
[docs] def subontology(self, minimal=False):
"""
Generates a sub-ontology based on associations
"""
return self.ontology.subontology(self.objects, minimal=minimal)
[docs] def associations(self, subject, object=None):
"""
Given a subject-object pair (e.g. gene id to ontology class id), return all association
objects that match.
"""
if object is None:
if self.associations_by_subj is not None:
return self.associations_by_subj[subject]
else:
return []
else:
if self.associations_by_subj_obj is not None:
return self.associations_by_subj_obj[(subject,object)]
else:
return []
# TODO: consider moving to other module
[docs] def enrichment_test(self, subjects=None, background=None, hypotheses=None, threshold=0.05, labels=False, direction='greater'):
"""
Performs term enrichment analysis.
Arguments
---------
subjects: string list
Sample set. Typically a gene ID list. These are assumed to have associations
background: string list
Background set. If not set, uses full set of known subject IDs in the association set
threshold: float
p values above this are filtered out
labels: boolean
if true, labels for enriched classes are included in result objects
direction: 'greater', 'less' or 'two-sided'
default is greater - i.e. enrichment test. Use 'less' for depletion test.
"""
if subjects is None:
subjects = []
subjects=set(subjects)
bg_count = {}
sample_count = {}
potential_hypotheses = set()
sample_size = len(subjects)
for s in subjects:
potential_hypotheses.update(self.inferred_types(s))
if hypotheses is None:
hypotheses = potential_hypotheses
else:
hypotheses = potential_hypotheses.intersection(hypotheses)
logger.info("Hypotheses: {}".format(hypotheses))
# get background counts
# TODO: consider doing this ahead of time
if background is None:
background = set(self.subjects)
else:
background = set(background)
# ensure background includes all subjects
background.update(subjects)
bg_size = len(background)
for c in hypotheses:
bg_count[c] = 0
sample_count[c] = 0
for s in background:
ancs = self.inferred_types(s)
for a in ancs.intersection(hypotheses):
bg_count[a] = bg_count[a]+1
for s in subjects:
for a in self.inferred_types(s):
if a in hypotheses:
sample_count[a] = sample_count[a]+1
hypotheses = [x for x in hypotheses if bg_count[x] > 1]
logger.info("Filtered hypotheses: {}".format(hypotheses))
num_hypotheses = len(hypotheses)
results = []
for cls in hypotheses:
# https://en.wikipedia.org/wiki/Fisher's_exact_test
#
# Cls NotCls RowTotal
# --- ------ ---
# study/sample [a, b] sample_size
# rest of ref [c, d] bg_size - sample_size
# --- ---
# nCls nNotCls
a = sample_count[cls]
b = sample_size - a
c = bg_count[cls] - a
d = (bg_size - bg_count[cls]) - b
#logger.debug("ABCD="+str((cls,a,b,c,d,sample_size)))
_, p_uncorrected = sp.stats.fisher_exact( [[a, b], [c, d]], direction)
p = p_uncorrected * num_hypotheses
if p>1.0:
p=1.0
#logger.debug("P={} uncorrected={}".format(p,p_uncorrected))
if p<threshold:
res = {'c':cls,'p':p,'p_uncorrected':p_uncorrected}
if labels:
res['n'] = self.ontology.label(cls)
results.append(res)
results = sorted(results, key=lambda x:x['p'])
return results
[docs] def jaccard_similarity(self,s1,s2):
"""
Calculate jaccard index of inferred associations of two subjects
|ancs(s1) /\ ancs(s2)|
---
|ancs(s1) \/ ancs(s2)|
"""
a1 = self.inferred_types(s1)
a2 = self.inferred_types(s2)
num_union = len(a1.union(a2))
if num_union == 0:
return 0.0
return len(a1.intersection(a2)) / num_union
[docs] def similarity_matrix(self, x_subjects=None, y_subjects=None, symmetric=False):
"""
Query for similarity matrix between groups of subjects
Return a list of intersection result objects with keys:
- x : term from x
- y : term from y
- c : count of intersection
- j : jaccard score
"""
if x_subjects is None:
x_subjects = []
if y_subjects is None:
y_subjects = []
xset = set(x_subjects)
yset = set(y_subjects)
zset = xset.union(yset)
# first built map of gene->termClosure.
# this could be calculated ahead of time for all g,
# but this may be space-expensive. TODO: benchmark
gmap={}
for z in zset:
gmap[z] = self.inferred_types(z)
ilist = []
for x in x_subjects:
for y in y_subjects:
if not symmetric or x<y:
shared = gmap[x].intersection(gmap[y])
union = gmap[x].union(gmap[y])
j = 0
if len(union)>0:
j = len(shared) / len(union)
ilist.append({'x':x,'y':y,'shared':shared, 'c':len(shared), 'j':j})
return self.intersectionlist_to_matrix(ilist, x_subjects, y_subjects)
class NamedEntity():
"""
E.g. a gene etc
"""
def __init__(self, id, label=None, taxon=None):
self.id=id
self.label=label
self.taxon=taxon
class AssociationSetMetadata():
"""
Information about how an association set is derived
"""
def __init__(self, id=None, taxon=None, evidence=None, subject_category=None, object_category=None):
self.id=id
self.taxon=taxon