def update_beliefs(self, corpus_id):
"""Return updated belief scores for a given corpus.
Parameters
----------
corpus_id : str
The ID of the corpus for which beliefs are to be updated.
Returns
-------
dict
A dictionary of belief scores with keys corresponding to Statement
UUIDs and values to new belief scores.
"""
corpus = self.get_corpus(corpus_id)
be = BeliefEngine(self.scorer)
stmts = list(corpus.statements.values())
be.set_prior_probs(stmts)
# Here we set beliefs based on actual curation
for uuid, correct in corpus.curations.items():
stmt = corpus.statements.get(uuid)
if stmt is None:
logger.warning('%s is not in the corpus.' % uuid)
continue
stmt.belief = correct
belief_dict = {st.uuid: st.belief for st in stmts}
return belief_dict
def update_beliefs(self, corpus_id):
"""Return updated belief scores for a given corpus.
Parameters
----------
corpus_id : str
The ID of the corpus for which beliefs are to be updated.
Returns
-------
dict
A dictionary of belief scores with keys corresponding to Statement
UUIDs and values to new belief scores.
"""
corpus = self.get_corpus(corpus_id)
be = BeliefEngine(self.scorer)
stmts = list(corpus.statements.values())
be.set_prior_probs(stmts)
# Here we set beliefs based on actual curation
for uuid, correct in corpus.curations.items():
stmt = corpus.statements.get(uuid)
if stmt is None:
logger.warning('%s is not in the corpus.' % uuid)
continue
stmt.belief = correct
belief_dict = {st.uuid: st.belief for st in stmts}
return belief_dict
def test_prior_prob_assertion():
be = BeliefEngine()
st = Phosphorylation(None, Agent('a'),
evidence=[ev1, deepcopy(ev1), ev2, ev3])
assert st.belief == 1
be.set_prior_probs([st])
assert st.belief == 1
def test_prior_prob_one():
be = BeliefEngine()
prob = 1 - (default_probs['rand']['reach'] +
default_probs['syst']['reach'])
st = Phosphorylation(None, Agent('a'), evidence=[ev1])
assert st.belief == 1
be.set_prior_probs([st])
assert st.belief == prob
def test_prior_prob_one():
be = BeliefEngine()
prob = 1 - (be.prior_probs['rand']['reach'] +
be.prior_probs['syst']['reach'])
st = Phosphorylation(None, Agent('a'), evidence=[ev1])
assert(st.belief == 1)
be.set_prior_probs([st])
assert(st.belief == prob)
def test_prior_prob_two_same():
be = BeliefEngine()
prob = 1 - (be.prior_probs['rand']['reach']**2 +
be.prior_probs['syst']['reach'])
st = Phosphorylation(None, Agent('a'), evidence=[ev1, ev1])
assert(st.belief == 1)
be.set_prior_probs([st])
assert(st.belief == prob)
def test_prior_prob_two_same():
be = BeliefEngine()
prob = 1 - (default_probs['rand']['reach']**2 +
default_probs['syst']['reach'])
st = Phosphorylation(None, Agent('a'), evidence=[ev1, ev1])
assert st.belief == 1
be.set_prior_probs([st])
assert st.belief == prob
def test_cycle():
st1 = Phosphorylation(Agent('B'), Agent('A1'))
st2 = Phosphorylation(None, Agent('A1'))
st1.supports = [st2]
st1.supported_by = [st2]
st2.supports = [st1]
st2.supported_by = [st1]
engine = BeliefEngine()
engine.set_hierarchy_probs([st1, st2])
def test_hierarchy_probs1():
be = BeliefEngine()
st1 = Phosphorylation(None, Agent('a'), evidence=[ev1])
st2 = Phosphorylation(None, Agent('b'), evidence=[ev2])
st2.supports = [st1]
st1.supported_by = [st2]
be.set_hierarchy_probs([st1, st2])
assert_close_enough(st1.belief, 1-0.35)
assert_close_enough(st2.belief, 1-0.35*0.35)
def test_hierarchy_probs1():
be = BeliefEngine()
st1 = Phosphorylation(None, Agent('a'), evidence=[ev1])
st2 = Phosphorylation(None, Agent('b'), evidence=[ev2])
st2.supports = [st1]
st1.supported_by = [st2]
be.set_hierarchy_probs([st1, st2])
assert_close_enough(st1.belief, 1 - 0.35)
assert_close_enough(st2.belief, 1 - 0.35 * 0.35)
def test_cycle():
st1 = Phosphorylation(Agent('B'), Agent('A1'))
st2 = Phosphorylation(None, Agent('A1'))
st1.supports = [st2]
st1.supported_by = [st2]
st2.supports = [st1]
st2.supported_by = [st1]
engine = BeliefEngine()
engine.set_hierarchy_probs([st1, st2])
def test_wm_scorer():
scorer = wm_scorer.get_eidos_scorer()
stmt = Influence(Concept('a'), Concept('b'),
evidence=[Evidence(source_api='eidos')])
# Make sure other sources are still in the map
assert 'hume' in scorer.prior_probs['rand']
assert 'biopax' in scorer.prior_probs['syst']
engine = BeliefEngine(scorer)
engine.set_prior_probs([stmt])
def test_evidence_random_noise_prior():
type_probs = {'biopax': 0.9, 'geneways': 0.2}
biopax_subtype_probs = {'reactome': 0.4, 'biogrid': 0.2}
geneways_subtype_probs = {'phosphorylate': 0.5, 'bind': 0.7}
subtype_probs = {
'biopax': biopax_subtype_probs,
'geneways': geneways_subtype_probs
}
ev_geneways_bind = Evidence(source_api='geneways',
source_id=0,
pmid=0,
text=None,
epistemics={},
annotations={'actiontype': 'bind'})
ev_biopax_reactome = Evidence(source_api='biopax',
source_id=0,
pmid=0,
text=None,
epistemics={},
annotations={'source_sub_id': 'reactome'})
ev_biopax_pid = Evidence(source_api='biopax',
source_id=0,
pmid=0,
text=None,
epistemics={},
annotations={'source_sub_id': 'pid'})
# Random noise prior for geneways bind evidence is the subtype prior,
# since we specified it
assert evidence_random_noise_prior(ev_geneways_bind, \
type_probs, subtype_probs) == 0.7
# Random noise prior for reactome biopax evidence is the subtype prior,
# since we specified it
assert evidence_random_noise_prior(ev_biopax_reactome, \
type_probs, subtype_probs) == 0.4
# Random noise prior for pid evidence is the subtype prior,
# since we specified it
assert evidence_random_noise_prior(ev_biopax_pid, type_probs,
subtype_probs) == 0.9
# Make sure this all still works when we go through the belief engine
statements = []
members = [Agent('a'), Agent('b')]
statements.append(Complex(members, evidence=ev_geneways_bind))
statements.append(Complex(members, evidence=ev_biopax_reactome))
statements.append(Complex(members, evidence=ev_biopax_pid))
p = {'rand': type_probs, 'syst': {'biopax': 0, 'geneways': 0}}
scorer = SimpleScorer(p, subtype_probs)
engine = BeliefEngine(scorer)
engine.set_prior_probs(statements)
assert statements[0].belief == 1 - 0.7
assert statements[1].belief == 1 - 0.4
assert statements[2].belief == 1 - 0.9
def test_wm_scorer():
scorer = wm_scorer.get_eidos_scorer()
stmt = Influence(Concept('a'),
Concept('b'),
evidence=[Evidence(source_api='eidos')])
# Make sure other sources are still in the map
assert 'hume' in scorer.prior_probs['rand']
assert 'biopax' in scorer.prior_probs['syst']
engine = BeliefEngine(scorer)
engine.set_prior_probs([stmt])
def test_prior_prob_one_two():
be = BeliefEngine()
prob = 1 - (default_probs['rand']['reach']**2 +
default_probs['syst']['reach']) * \
(default_probs['rand']['trips'] +
default_probs['syst']['trips'])
st = Phosphorylation(None, Agent('a'), evidence=[ev1, deepcopy(ev1), ev2])
assert st.belief == 1
be.set_prior_probs([st])
assert st.belief == prob
def test_hierarchy_probs1():
be = BeliefEngine()
st1 = Phosphorylation(None, Agent('a'), evidence=[ev1])
st2 = Phosphorylation(None, Agent('b'), evidence=[ev2])
st2.supports = [st1]
st1.supported_by = [st2]
st1.belief = 0.5
st2.belief = 0.8
be.set_hierarchy_probs([st1, st2])
assert(st1.belief == 0.5)
assert(st2.belief == 0.9)
def test_evidence_random_noise_prior():
type_probs = {'biopax': 0.9, 'geneways': 0.2}
biopax_subtype_probs = {
'reactome': 0.4,
'biogrid': 0.2}
geneways_subtype_probs = {
'phosphorylate': 0.5,
'bind': 0.7}
subtype_probs = {'biopax': biopax_subtype_probs,
'geneways': geneways_subtype_probs}
ev_geneways_bind = Evidence(source_api='geneways', source_id=0,
pmid=0, text=None, epistemics={},
annotations={'actiontype': 'bind'})
ev_biopax_reactome = Evidence(source_api='biopax', source_id=0,
pmid=0, text=None, epistemics={},
annotations={'source_sub_id': 'reactome'})
ev_biopax_pid = Evidence(source_api='biopax', source_id=0,
pmid=0, text=None, epistemics={},
annotations={'source_sub_id': 'pid'})
# Random noise prior for geneways bind evidence is the subtype prior,
# since we specified it
assert evidence_random_noise_prior(ev_geneways_bind, \
type_probs, subtype_probs) == 0.7
# Random noise prior for reactome biopax evidence is the subtype prior,
# since we specified it
assert evidence_random_noise_prior(ev_biopax_reactome, \
type_probs, subtype_probs) == 0.4
# Random noise prior for pid evidence is the subtype prior,
# since we specified it
assert evidence_random_noise_prior(ev_biopax_pid,
type_probs, subtype_probs) == 0.9
# Make sure this all still works when we go through the belief engine
statements = []
members = [Agent('a'), Agent('b')]
statements.append(Complex(members, evidence=ev_geneways_bind))
statements.append(Complex(members, evidence=ev_biopax_reactome))
statements.append(Complex(members, evidence=ev_biopax_pid))
p = {'rand': type_probs, 'syst': {'biopax': 0, 'geneways': 0}}
scorer = SimpleScorer(p, subtype_probs)
engine = BeliefEngine(scorer)
engine.set_prior_probs(statements)
assert statements[0].belief == 1 - 0.7
assert statements[1].belief == 1 - 0.4
assert statements[2].belief == 1 - 0.9
def test_belief_calc_up_to_prior():
be = BeliefEngine()
test_stmts = [
MockStatement(1, [MockEvidence('sparser'), MockEvidence('reach')]),
MockStatement(2, MockEvidence('biopax')),
MockStatement(3, MockEvidence('signor')),
MockStatement(4, MockEvidence('biogrid')),
MockStatement(5, MockEvidence('bel')),
MockStatement(6, [MockEvidence('phosphosite'), MockEvidence('trips')]),
]
be.set_prior_probs(test_stmts)
results = {s.matches_key(): s.belief for s in test_stmts}
print(results)
assert len(results) == len(test_stmts), (len(results), len(test_stmts))
assert all([0 < b < 1 for b in results.values()]), 'Beliefs out of range.'
def setup_belief(include_more_specific=False):
# Make a model
lr = LogisticRegression()
# Get all the sources
source_list = CountsScorer.get_all_sources(test_stmts_cur)
cs = CountsScorer(lr,
source_list,
include_more_specific=include_more_specific)
# Train on curated stmt data
if include_more_specific:
extra_evidence = [[
ev for supp in stmt.supports for ev in supp.evidence
] for stmt in test_stmts_cur]
else:
extra_evidence = None
# Fit with extra evidence, if any
cs.fit(test_stmts_cur, y_arr_stmts_cur, extra_evidence)
# Run predictions on test statements without extra evidence to get prior
# probs
probs = cs.predict_proba(test_stmts_cur)[:, 1]
# Now check if we get these same beliefs set on the statements when we
# run with the belief engine:
# Get scorer and belief engine instances for trained model
be = BeliefEngine(scorer=cs)
# Make a shallow copy of the test stmts so that we don't change beliefs
# of the global instances as a side-effect of this test
test_stmts_copy = copy(test_stmts_cur)
return be, test_stmts_copy, probs
def test_hierarchy_probs4():
be = BeliefEngine()
st1 = Phosphorylation(None, Agent('a'), evidence=[ev1])
st2 = Phosphorylation(None, Agent('b'), evidence=[ev2])
st3 = Phosphorylation(None, Agent('c'), evidence=[ev1])
st4 = Phosphorylation(None, Agent('d'), evidence=[ev1])
st4.supports = [st1, st2, st3]
st3.supports = [st1]
st2.supports = [st1]
st1.supported_by = [st2, st3, st4]
st2.supported_by = [st4]
st3.supported_by = [st4]
be.set_hierarchy_probs([st1, st2, st3, st4])
assert_close_enough(st1.belief, 1 - 0.35)
assert_close_enough(st2.belief, 1 - 0.35 * 0.35)
assert_close_enough(st3.belief, 1 - (0.05 + 0.3 * 0.3))
assert_close_enough(st4.belief, 1 - 0.35 * (0.05 + 0.3 * 0.3 * 0.3))
def test_hierarchy_probs4():
be = BeliefEngine()
st1 = Phosphorylation(None, Agent('a'), evidence=[ev1])
st2 = Phosphorylation(None, Agent('b'), evidence=[ev2])
st3 = Phosphorylation(None, Agent('c'), evidence=[ev1])
st4 = Phosphorylation(None, Agent('d'), evidence=[ev1])
st4.supports = [st1, st2, st3]
st3.supports = [st1]
st2.supports = [st1]
st1.supported_by = [st2, st3, st4]
st2.supported_by = [st4]
st3.supported_by = [st4]
be.set_hierarchy_probs([st1, st2, st3, st4])
assert_close_enough(st1.belief, 1-0.35)
assert_close_enough(st2.belief, 1-0.35*0.35)
assert_close_enough(st3.belief, 1-(0.05 + 0.3*0.3))
assert_close_enough(st4.belief, 1-0.35*(0.05 + 0.3*0.3*0.3))
def test_default_probs_override():
"""Make sure default probs are overriden by constructor argument."""
be = BeliefEngine(prior_probs={'rand': {'assertion': 0.5}})
for err_type in ('rand', 'syst'):
for k, v in be.prior_probs[err_type].items():
if err_type == 'rand' and k == 'assertion':
assert v == 0.5
else:
assert default_probs[err_type][k] == v
def test_negative_evidence():
prior_probs = {'rand': {'new_source': 0.1},
'syst': {'new_source': 0.05}}
getev = lambda x: Evidence(source_api='new_source',
epistemics={'negated': x})
evs1 = [getev(x) for x in [True, True, False]]
evs2 = [getev(x) for x in [False, False, False]]
evs3 = [getev(x) for x in [True, True, True]]
stmts = [Phosphorylation(None, Agent('a'), evidence=e)
for e in [evs1, evs2, evs3]]
scorer = SimpleScorer(prior_probs)
engine = BeliefEngine(scorer)
engine.set_prior_probs(stmts)
pr = prior_probs['rand']['new_source']
ps = prior_probs['syst']['new_source']
assert_close_enough(stmts[0].belief, ((1-pr)-ps)*(1-((1-pr*pr)-ps)))
assert_close_enough(stmts[1].belief, (1-pr*pr*pr)-ps)
assert stmts[2].belief == 0
def test_negative_evidence():
prior_probs = {'rand': {'new_source': 0.1},
'syst': {'new_source': 0.05}}
getev = lambda x: Evidence(source_api='new_source',
epistemics={'negated': x})
evs1 = [getev(x) for x in [True, True, False]]
evs2 = [getev(x) for x in [False, False, False]]
evs3 = [getev(x) for x in [True, True, True]]
stmts = [Phosphorylation(None, Agent('a'), evidence=e)
for e in [evs1, evs2, evs3]]
scorer = SimpleScorer(prior_probs)
engine = BeliefEngine(scorer)
engine.set_prior_probs(stmts)
pr = prior_probs['rand']['new_source']
ps = prior_probs['syst']['new_source']
assert_close_enough(stmts[0].belief, ((1-pr)-ps)*(1-((1-pr*pr)-ps)))
assert_close_enough(stmts[1].belief, (1-pr*pr*pr)-ps)
assert stmts[2].belief == 0
def test_hierarchy_probs4():
be = BeliefEngine()
st1 = Phosphorylation(None, Agent('a'), evidence=[ev1])
st2 = Phosphorylation(None, Agent('b'), evidence=[ev2])
st3 = Phosphorylation(None, Agent('c'), evidence=[ev3])
st4 = Phosphorylation(None, Agent('d'), evidence=[ev1])
st4.supports = [st1, st2, st3]
st3.supports = [st1]
st2.supports = [st1]
st1.supported_by = [st2, st3, st4]
st2.supported_by = [st4]
st3.supported_by = [st4]
st1.belief = 0.5
st2.belief = 0.8
st3.belief = 0.2
st4.belief = 0.6
be.set_hierarchy_probs([st1, st2, st3])
assert(st1.belief == 0.5)
assert(st2.belief == 0.9)
assert(st3.belief == 0.6)
assert(st4.belief == 0.968)
def calculate_belief(stmts):
scorer = SimpleScorer(subtype_probs={
'biopax': {'pc11': 0.2, 'phosphosite': 0.01},
})
be = BeliefEngine(scorer=scorer)
be.set_prior_probs(stmts)
be.set_hierarchy_probs(stmts)
return {str(s.get_hash()): s.belief for s in stmts}
def test_default_probs_extend():
"""Make sure default probs are extended by constructor argument."""
prior_probs = {'rand': {'new_source': 0.1}, 'syst': {'new_source': 0.05}}
scorer = SimpleScorer(prior_probs)
be = BeliefEngine(scorer)
for err_type in ('rand', 'syst'):
assert 'new_source' in scorer.prior_probs[err_type]
for k, v in scorer.prior_probs[err_type].items():
if err_type == 'rand' and k == 'new_source':
assert v == 0.1
elif err_type == 'syst' and k == 'new_source':
assert v == 0.05
else:
assert default_probs[err_type][k] == v
def run_preassembly(stmts_in, **kwargs):
"""Run preassembly on a list of statements.
Parameters
----------
stmts_in : list[indra.statements.Statement]
A list of statements to preassemble.
return_toplevel : Optional[bool]
If True, only the top-level statements are returned. If False,
all statements are returned irrespective of level of specificity.
Default: True
poolsize : Optional[int]
The number of worker processes to use to parallelize the
comparisons performed by the function. If None (default), no
parallelization is performed. NOTE: Parallelization is only
available on Python 3.4 and above.
size_cutoff : Optional[int]
Groups with size_cutoff or more statements are sent to worker
processes, while smaller groups are compared in the parent process.
Default value is 100. Not relevant when parallelization is not
used.
save : Optional[str]
The name of a pickle file to save the results (stmts_out) into.
save_unique : Optional[str]
The name of a pickle file to save the unique statements into.
Returns
-------
stmts_out : list[indra.statements.Statement]
A list of preassembled top-level statements.
"""
dump_pkl_unique = kwargs.get('save_unique')
be = BeliefEngine()
pa = Preassembler(hierarchies, stmts_in)
run_preassembly_duplicate(pa, be, save=dump_pkl_unique)
dump_pkl = kwargs.get('save')
return_toplevel = kwargs.get('return_toplevel', True)
poolsize = kwargs.get('poolsize', None)
size_cutoff = kwargs.get('size_cutoff', 100)
options = {
'save': dump_pkl,
'return_toplevel': return_toplevel,
'poolsize': poolsize,
'size_cutoff': size_cutoff
}
stmts_out = run_preassembly_related(pa, be, **options)
return stmts_out
def setup_belief():
# Make a model
lr = LogisticRegression()
# Get all the sources
source_list = CountsScorer.get_all_sources(test_stmts_cur)
cs = CountsScorer(lr, source_list)
# Train on curated stmt data
cs.fit(test_stmts_cur, y_arr_stmts_cur)
# Run predictions on test statements
probs = cs.predict_proba(test_stmts_cur)[:, 1]
# Now check if we get these same beliefs set on the statements when we
# run with the belief engine:
# Get scorer and belief engine instances for trained model
be = BeliefEngine(scorer=cs)
# Make a shallow copy of the test stmts so that we don't change beliefs
# of the global instances as a side-effect of this test
test_stmts_copy = copy(test_stmts_cur)
return be, test_stmts_copy, probs
def test_belief_calc_up_to_hierarchy():
be = BeliefEngine()
test_stmts = [
MockStatement(1, [MockEvidence('sparser'),
MockEvidence('reach')]),
MockStatement(2, MockEvidence('biopax')),
MockStatement(3, MockEvidence('signor')),
MockStatement(4, MockEvidence('biogrid')),
MockStatement(5, MockEvidence('bel')),
MockStatement(6, [MockEvidence('phosphosite'),
MockEvidence('trips')]),
]
be.set_prior_probs(test_stmts)
init_results = {s.matches_key(): s.belief for s in test_stmts}
print(init_results)
supp_links = [(1, 2), (1, 3), (2, 3), (1, 5), (4, 3)]
populate_support(test_stmts, supp_links)
be.set_hierarchy_probs(test_stmts)
results = {s.matches_key(): s.belief for s in test_stmts}
print(results)
# Test a couple very simple properties.
assert len(results) == len(test_stmts), (len(results), len(test_stmts))
assert all([0 < b < 1 for b in results.values()]), 'Beliefs out of range.'
# Test the change from the initial.
all_deltas_correct = True
deltas_dict = {}
for s in test_stmts:
h = s.matches_key()
b = s.belief
# Get results
res = {'actual': b - init_results[h]}
# Define expectations.
if s.supports:
res['expected'] = 'increase'
if res['actual'] <= 0:
all_deltas_correct = False
else:
res['expected'] = 'no change'
if res['actual'] != 0:
all_deltas_correct = False
deltas_dict[h] = res
assert all_deltas_correct, deltas_dict
def run_preassembly(stmts_in, **kwargs):
"""Run preassembly on a list of statements.
Parameters
----------
stmts_in : list[indra.statements.Statement]
A list of statements to preassemble.
return_toplevel : Optional[bool]
If True, only the top-level statements are returned. If False,
all statements are returned irrespective of level of specificity.
Default: True
save : Optional[str]
The name of a pickle file to save the results (stmts_out) into.
save_unique : Optional[str]
The name of a pickle file to save the unique statements into.
Returns
-------
stmts_out : list[indra.statements.Statement]
A list of preassembled top-level statements.
"""
dump_pkl = kwargs.get('save')
dump_pkl_unique = kwargs.get('save_unique')
be = BeliefEngine()
pa = Preassembler(hierarchies, stmts_in)
options = {'save': dump_pkl_unique}
run_preassembly_duplicate(pa, be, **options)
return_toplevel = kwargs.get('return_toplevel', True)
options = {'save': dump_pkl, 'return_toplevel': return_toplevel}
start = time.time()
stmts_out = run_preassembly_related(pa, be, **options)
end = time.time()
elapsed = end - start
logger.debug("Time elapsed, run_preassembly_related: %s" % elapsed)
return stmts_out
def run_preassembly(statements, hierarchies):
print('%d total statements' % len(statements))
# Filter to grounded only
statements = ac.filter_grounded_only(statements, score_threshold=0.4)
# Make a Preassembler with the Eidos and TRIPS ontology
pa = Preassembler(hierarchies, statements)
# Make a BeliefEngine and run combine duplicates
be = BeliefEngine()
unique_stmts = pa.combine_duplicates()
print('%d unique statements' % len(unique_stmts))
be.set_prior_probs(unique_stmts)
# Run combine related
related_stmts = pa.combine_related(return_toplevel=False)
be.set_hierarchy_probs(related_stmts)
# Filter to top-level Statements
top_stmts = ac.filter_top_level(related_stmts)
print('%d top-level statements' % len(top_stmts))
return top_stmts
def run_preassembly(statements, hierarchies):
print('%d total statements' % len(statements))
# Filter to grounded only
statements = map_onto(statements)
ac.dump_statements(statements, 'pi_mtg_demo_unfiltered.pkl')
statements = ac.filter_grounded_only(statements, score_threshold=0.7)
#statements = ac.filter_by_db_refs(statements, 'UN',
# ['conflict', 'food_security', 'precipitation'], policy='one',
# match_suffix=True)
statements = ac.filter_by_db_refs(
statements,
'UN', [
'conflict', 'food_security', 'flooding', 'food_production',
'human_migration', 'drought', 'food_availability', 'market',
'food_insecurity'
],
policy='all',
match_suffix=True)
assume_polarity(statements)
statements = filter_has_polarity(statements)
# Make a Preassembler with the Eidos and TRIPS ontology
pa = Preassembler(hierarchies, statements)
# Make a BeliefEngine and run combine duplicates
be = BeliefEngine()
unique_stmts = pa.combine_duplicates()
print('%d unique statements' % len(unique_stmts))
be.set_prior_probs(unique_stmts)
# Run combine related
related_stmts = pa.combine_related(return_toplevel=False)
be.set_hierarchy_probs(related_stmts)
#related_stmts = ac.filter_belief(related_stmts, 0.8)
# Filter to top-level Statements
top_stmts = ac.filter_top_level(related_stmts)
pa.stmts = top_stmts
print('%d top-level statements' % len(top_stmts))
conflicts = pa.find_contradicts()
top_stmts = remove_contradicts(top_stmts, conflicts)
ac.dump_statements(top_stmts, 'pi_mtg_demo.pkl')
return top_stmts
def run_assembly(stmts, folder, pmcid, background_assertions=None):
'''Run assembly on a list of statements, for a given PMCID.'''
# Folder for index card output (scored submission)
indexcard_prefix = folder + '/index_cards/' + pmcid
# Folder for other outputs (for analysis, debugging)
otherout_prefix = folder + '/other_outputs/' + pmcid
# Do grounding mapping here
# Load the TRIPS-specific grounding map and add to the default
# (REACH-oriented) grounding map:
trips_gm = load_grounding_map('trips_grounding_map.csv')
default_grounding_map.update(trips_gm)
gm = GroundingMapper(default_grounding_map)
mapped_agent_stmts = gm.map_agents(stmts)
renamed_agent_stmts = gm.rename_agents(mapped_agent_stmts)
# Filter for grounding
grounded_stmts = []
for st in renamed_agent_stmts:
if all([is_protein_or_chemical(a) for a in st.agent_list()]):
grounded_stmts.append(st)
# Instantiate the Preassembler
pa = Preassembler(hierarchies)
pa.add_statements(grounded_stmts)
print('== %s ====================' % pmcid)
print('%d statements collected in total.' % len(pa.stmts))
# Combine duplicates
unique_stmts = pa.combine_duplicates()
print('%d statements after combining duplicates.' % len(unique_stmts))
# Run BeliefEngine on unique statements
epe = BeliefEngine()
epe.set_prior_probs(pa.unique_stmts)
# Build statement hierarchy
related_stmts = pa.combine_related()
# Run BeliefEngine on hierarchy
epe.set_hierarchy_probs(related_stmts)
print('%d statements after combining related.' % len(related_stmts))
# Instantiate the mechanism linker
# Link statements
linked_stmts = MechLinker.infer_active_forms(related_stmts)
linked_stmts += MechLinker.infer_modifications(related_stmts)
linked_stmts += MechLinker.infer_activations(related_stmts)
# Run BeliefEngine on linked statements
epe.set_linked_probs(linked_stmts)
# Print linked statements for debugging purposes
print('Linked\n=====')
for ls in linked_stmts:
print(ls.inferred_stmt.belief, ls.inferred_stmt)
print('=============')
# Combine all statements including linked ones
all_statements = related_stmts + [ls.inferred_stmt for ls in linked_stmts]
# Instantiate a new preassembler
pa = Preassembler(hierarchies, all_statements)
# Build hierarchy again
pa.combine_duplicates()
# Choose the top-level statements
related_stmts = pa.combine_related()
# Remove top-level statements that came only from the prior
if background_assertions is not None:
nonbg_stmts = [stmt for stmt in related_stmts
if stmt not in background_assertions]
else:
nonbg_stmts = related_stmts
# Dump top-level statements in a pickle
with open(otherout_prefix + '.pkl', 'wb') as fh:
pickle.dump(nonbg_stmts, fh)
# Flatten evidence for statements
flattened_evidence_stmts = flatten_evidence(nonbg_stmts)
# Start a card counter
card_counter = 1
# We don't limit the number of cards reported in this round
card_lim = float('inf')
top_stmts = []
###############################################
# The belief cutoff for statements
belief_cutoff = 0.3
###############################################
# Sort by amount of evidence
for st in sorted(flattened_evidence_stmts,
key=lambda x: x.belief, reverse=True):
if st.belief >= belief_cutoff:
print(st.belief, st)
if st.belief < belief_cutoff:
print('SKIP', st.belief, st)
# If it's background knowledge, we skip the statement
if is_background_knowledge(st):
print('This statement is background knowledge - skipping.')
continue
# Assemble IndexCards
ia = IndexCardAssembler([st], pmc_override=pmcid)
ia.make_model()
# If the index card was actually made
# (not all statements can be assembled into index cards to
# this is often not the case)
if ia.cards:
# Save the index card json
ia.save_model(indexcard_prefix + '-%d.json' % card_counter)
card_counter += 1
top_stmts.append(st)
if card_counter > card_lim:
break
# Print the English-assembled model for debugging purposes
ea = EnglishAssembler(top_stmts)
print('=======================')
print(ea.make_model().encode('utf-8'))
print('=======================')
# Print the statement graph
graph = render_stmt_graph(nonbg_stmts)
graph.draw(otherout_prefix + '_graph.pdf', prog='dot')
# Print statement diagnostics
print_stmts(pa.stmts, otherout_prefix + '_statements.tsv')
print_stmts(related_stmts, otherout_prefix + '_related_statements.tsv')
if __name__ == '__main__':
if len(sys.argv) < 3:
logger.error('Usage: assemble_corpus.py <pickle_file> <output_folder>')
sys.exit()
stmts_fname = sys.argv[1]
out_folder = sys.argv[2]
stmts = load_statements(stmts_fname)
logger.info('All statements: %d' % len(stmts))
cache_pkl = os.path.join(out_folder, 'mapped_stmts.pkl')
options = {'save': cache_pkl, 'do_rename': True}
stmts = map_grounding(stmts, **options)
cache_pkl = os.path.join(out_folder, 'sequence_valid_stmts.pkl')
options = {'save': cache_pkl}
mapped_stmts = map_sequence(stmts, **options)
be = BeliefEngine()
pa = Preassembler(hierarchies, mapped_stmts)
cache_pkl = os.path.join(out_folder, 'unique_stmts.pkl')
options = {'save': cache_pkl}
unique_stmts = run_preassembly_duplicate(pa, be, **options)
cache_pkl = os.path.join(out_folder, 'top_stmts.pkl')
options = {'save': cache_pkl}
stmts = run_preassembly_related(pa, be, **options)
def run_assembly(stmts, folder, pmcid, background_assertions=None):
'''Run assembly on a list of statements, for a given PMCID.'''
# Folder for index card output (scored submission)
indexcard_prefix = folder + '/index_cards/' + pmcid
# Folder for other outputs (for analysis, debugging)
otherout_prefix = folder + '/other_outputs/' + pmcid
# Do grounding mapping here
# Load the TRIPS-specific grounding map and add to the default
# (REACH-oriented) grounding map:
trips_gm = load_grounding_map('trips_grounding_map.csv')
default_grounding_map.update(trips_gm)
gm = GroundingMapper(default_grounding_map)
mapped_agent_stmts = gm.map_agents(stmts)
renamed_agent_stmts = gm.rename_agents(mapped_agent_stmts)
# Filter for grounding
grounded_stmts = []
for st in renamed_agent_stmts:
if all([is_protein_or_chemical(a) for a in st.agent_list()]):
grounded_stmts.append(st)
# Instantiate the Preassembler
pa = Preassembler(hierarchies)
pa.add_statements(grounded_stmts)
print('== %s ====================' % pmcid)
print('%d statements collected in total.' % len(pa.stmts))
# Combine duplicates
unique_stmts = pa.combine_duplicates()
print('%d statements after combining duplicates.' % len(unique_stmts))
# Run BeliefEngine on unique statements
epe = BeliefEngine()
epe.set_prior_probs(pa.unique_stmts)
# Build statement hierarchy
related_stmts = pa.combine_related()
# Run BeliefEngine on hierarchy
epe.set_hierarchy_probs(related_stmts)
print('%d statements after combining related.' % len(related_stmts))
# Instantiate the mechanism linker
ml = MechLinker(related_stmts)
# Link statements
linked_stmts = ml.link_statements()
# Run BeliefEngine on linked statements
epe.set_linked_probs(linked_stmts)
# Print linked statements for debugging purposes
print('Linked\n=====')
for ls in linked_stmts:
print(ls.inferred_stmt.belief, ls.inferred_stmt)
print('=============')
# Combine all statements including linked ones
all_statements = ml.statements + [ls.inferred_stmt for ls in linked_stmts]
# Instantiate a new preassembler
pa = Preassembler(hierarchies, all_statements)
# Build hierarchy again
pa.combine_duplicates()
# Choose the top-level statements
related_stmts = pa.combine_related()
# Remove top-level statements that came only from the prior
if background_assertions is not None:
nonbg_stmts = [
stmt for stmt in related_stmts if stmt not in background_assertions
]
else:
nonbg_stmts = related_stmts
# Dump top-level statements in a pickle
with open(otherout_prefix + '.pkl', 'wb') as fh:
pickle.dump(nonbg_stmts, fh, protocol=2)
# Flatten evidence for statements
flattened_evidence_stmts = flatten_evidence(nonbg_stmts)
# Start a card counter
card_counter = 1
# We don't limit the number of cards reported in this round
card_lim = float('inf')
top_stmts = []
###############################################
# The belief cutoff for statements
belief_cutoff = 0.3
###############################################
# Sort by amount of evidence
for st in sorted(flattened_evidence_stmts,
key=lambda x: x.belief,
reverse=True):
if st.belief >= belief_cutoff:
print(st.belief, st)
if st.belief < belief_cutoff:
print('SKIP', st.belief, st)
# If it's background knowledge, we skip the statement
if is_background_knowledge(st):
print('This statement is background knowledge - skipping.')
continue
# Assemble IndexCards
ia = IndexCardAssembler([st], pmc_override=pmcid)
ia.make_model()
# If the index card was actually made
# (not all statements can be assembled into index cards to
# this is often not the case)
if ia.cards:
# Save the index card json
ia.save_model(indexcard_prefix + '-%d.json' % card_counter)
card_counter += 1
top_stmts.append(st)
if card_counter > card_lim:
break
# Print the English-assembled model for debugging purposes
ea = EnglishAssembler(top_stmts)
print('=======================')
print(ea.make_model())
print('=======================')
# Print the statement graph
graph = render_stmt_graph(nonbg_stmts)
graph.draw(otherout_prefix + '_graph.pdf', prog='dot')
# Print statement diagnostics
print_stmts(pa.stmts, otherout_prefix + '_statements.tsv')
print_stmts(related_stmts, otherout_prefix + '_related_statements.tsv')
def test_prior_prob_assertion():
be = BeliefEngine()
st = Phosphorylation(None, Agent('a'), evidence=[ev1, ev1, ev2, ev3])
assert(st.belief == 1)
be.set_prior_probs([st])
assert(st.belief == 1)
def calculate_belief(stmts):
be = BeliefEngine()
be.set_prior_probs(stmts)
be.set_hierarchy_probs(stmts)
return {s.matches_key(): s.belief for s in stmts}
def test_check_prior_probs():
be = BeliefEngine()
st = Phosphorylation(None, Agent('ERK'),
evidence=[Evidence(source_api='xxx')])
be.set_prior_probs([st])
def preassemble(self, filters=None):
"""Preassemble the Statements collected in the model.
Use INDRA's GroundingMapper, Preassembler and BeliefEngine
on the IncrementalModel and save the unique statements and
the top level statements in class attributes.
Currently the following filter options are implemented:
- grounding: require that all Agents in statements are grounded
- model_one: require that at least one Agent is in the incremental model
- model_all: require that all Agents are in the incremental model
- prior_one: require that at least one Agent is in the prior model
- prior_all: require that all Agents are in the prior model
Note that model_one -> prior_all are increasingly more restrictive
options.
Parameters
----------
filters : Optional[list[str]]
A list of filter options to apply when choosing the statements.
See description above for more details. Default: None
"""
stmts = self.get_statements()
logger.info("%d raw Statements in total" % len(stmts))
# Fix grounding
logger.info("Running grounding map")
twg = gm.agent_texts_with_grounding(stmts)
prot_map = gm.protein_map_from_twg(twg)
gm.default_grounding_map.update(prot_map)
gmap = gm.GroundingMapper(gm.default_grounding_map)
stmts = gmap.map_agents(stmts, do_rename=True)
logger.info("%d Statements after grounding map" % len(stmts))
# Fix sites
sm = SiteMapper(default_site_map)
stmts, _ = sm.map_sites(stmts)
logger.info("%d Statements with valid sequence" % len(stmts))
if filters:
if "grounding" in filters:
# Filter out ungrounded statements
logger.info("Running grounding filter")
stmts = self._relevance_filter(stmts, ["grounding"])
logger.info("%s Statements after filter" % len(stmts))
if "human_only" in filters:
# Filter out non-human proteins
logger.info("Running non-human protein filter")
stmts = self._relevance_filter(stmts, ["human_only"])
logger.info("%s Statements after filter" % len(stmts))
for rel_key in ("prior_one", "model_one", "prior_all", "model_all"):
if rel_key in filters:
logger.info("Running %s relevance filter" % rel_key)
stmts = self._relevance_filter(stmts, [rel_key])
logger.info("%s Statements after filter" % len(stmts))
# Combine duplicates
logger.info("Preassembling %d Statements" % len(stmts))
pa = Preassembler(hierarchies, stmts)
self.unique_stmts = pa.combine_duplicates()
logger.info("%d unique Statements" % len(self.unique_stmts))
# Run BeliefEngine on unique statements
be = BeliefEngine()
be.set_prior_probs(self.unique_stmts)
# Build statement hierarchy
self.unique_stmts = pa.combine_related(return_toplevel=False)
self.toplevel_stmts = [st for st in self.unique_stmts if not st.supports]
logger.info("%d top-level Statements" % len(self.toplevel_stmts))
# Run BeliefEngine on hierarchy
be.set_hierarchy_probs(self.unique_stmts)
generate belief scores for each of the statements and output a new pickle
file containing a list of statements with beliefs. Script throws away the
dictionaries keys. Input statements should contain evidence, otherwise
their belief scores will be set to 0."""
parser = ArgumentParser(description=description, epilog=epilog)
parser.add_argument("-d",
action="store_true",
help="set if input is a dictionary of statements")
help_text = ("path to a pickle file containing a list or dict of"
"statements.")
parser.add_argument("infile", help=help_text)
args = parser.parse_args()
infile = args.infile
filename, file_extension = path.splitext(infile)
outfile = filename + "_with_beliefs" + file_extension
with open(args.infile, 'rb') as f:
stmts = pickle.load(f)
if args.d:
stmts = [stmt for _, stmt in stmts.items()]
# get belief scores
for stmt in stmts:
stmt.belief = 1
be = BeliefEngine()
be.set_prior_probs(stmts)
# using pickle instead of assemble_corpus to avoid printing logging
with open(outfile, 'wb') as f:
pickle.dump(stmts, f)
def test_check_prior_probs():
be = BeliefEngine()
st = Phosphorylation(None,
Agent('ERK'),
evidence=[Evidence(source_api='xxx')])
be.set_prior_probs([st])
def test_default_probs():
"""Make sure default probs are set with empty constructor."""
be = BeliefEngine()
for err_type in ('rand', 'syst'):
for k, v in default_probs[err_type].items():
assert default_probs[err_type][k] == v