def __init__(self, **kwargs):
"""Initializes TemplateFreeNeuralNetScorer.
Args:
**kwargs: Unused.
"""
from makeit.synthetic.evaluation.rexgen_direct.predict import TFFP
self.model = TFFP()
def __init__(self, **kwargs):
from makeit.synthetic.evaluation.rexgen_direct.predict import TFFP
self.model = TFFP()
class TemplateFreeNeuralNetScorer(Scorer):
"""Template-free neural net evaluator.
Attributes:
model (makeit.synthetic.evaluation.rexgen_direct.predict.TFFP):
Template-free forward predictor.
"""
def __init__(self, **kwargs):
"""Initializes TemplateFreeNeuralNetScorer.
Args:
**kwargs: Unused.
"""
from makeit.synthetic.evaluation.rexgen_direct.predict import TFFP
self.model = TFFP()
def evaluate(self,
reactants_smiles,
contexts=[(20, '', '', '', '', '')],
**kwargs):
"""Evaluates possible reaction outcomes for given reactants.
Evaluation does not use context, but left as dummy pos var.
Args:
reactants_smiles (list of str??): SMILES string of reactants.
contexts (list, optional): Unused.
(default: {[(20,'','','','','')]})
**kwargs: Additional optional parameters.
Returns:
list: Predicted reaction outcomes.
"""
all_outcomes = []
for (T1, slvt1, rgt1, cat1, t1, y1) in contexts:
this_reactants_smiles = reactants_smiles
if slvt1:
this_reactants_smiles += '.' + slvt1
if rgt1:
this_reactants_smiles += '.' + rgt1
if cat1:
this_reactants_smiles += '.' + cat1
_, outcomes = self.model.predict(this_reactants_smiles,
top_n=kwargs.get('top_n', 1e10))
if not outcomes:
all_outcomes.append([{
'rank': 1,
'outcome': {
'smiles': '',
'template_ids': [],
'num_examples': 0
},
'score': 0,
'prob': 0,
}])
continue
outcomes_to_ret = {}
reactants_smiles_split = this_reactants_smiles.split('.')
for outcome in outcomes:
smiles_list = set(outcome['smiles'].split('.'))
# Canonicalize
smiles_canonical = set()
for smi in smiles_list:
mol = Chem.MolFromSmiles(smi)
if not mol:
MyLogger.print_and_log(
'Template free evaluator could not reparse product {}'
.format('.'.join(smiles_list)),
template_free_scorer_loc, 1)
continue
smiles_canonical.add(Chem.MolToSmiles(mol))
# Remove unreacted frags
smiles_canonical = smiles_canonical - set(
reactants_smiles_split)
if not smiles_canonical:
continue # no reaction?
smiles = max(
smiles_canonical, key=len
) # NOTE: this is not great...byproducts may be longer
if not smiles:
continue
if smiles in outcomes_to_ret:
outcomes_to_ret[smiles]['rank'] = min(
outcomes_to_ret[smiles]['rank'], outcome['rank'])
outcomes_to_ret[smiles]['score'] = np.log(
np.exp(outcomes_to_ret[smiles]['score']) +
np.exp(outcome['score']))
outcomes_to_ret[smiles]['prob'] += outcome['prob']
else:
# Append outcome information
outcomes_to_ret[smiles] = {
'rank':
outcome['rank'],
'outcome': {
'smiles': smiles,
# 'smiles_list': smiles_list,
'template_ids': [],
'num_examples': 0,
},
'score':
float(outcome['score']),
'prob':
float(outcome['prob']),
'mol_wt':
float(Descriptors.MolWt(Chem.MolFromSmiles(smiles)))
}
# Renormalize and re-rank
outcomes = sorted(outcomes_to_ret.values(),
key=lambda x: x['prob'],
reverse=True)
total_prob = sum([outcome['prob'] for outcome in outcomes])
for i, outcome in enumerate(outcomes):
outcomes[i]['rank'] = i + 1
outcomes[i]['prob'] = outcome['prob'] / total_prob
all_outcomes.append(outcomes)
# Return in lists as if we received a list of contexts
return all_outcomes
class WLN_AtomMapper:
"""Template-free neural net evaluator.
Attributes:
model (makeit.synthetic.evaluation.rexgen_direct.predict.TFFP):
Template-free forward predictor.
"""
def __init__(self, **kwargs):
"""Initializes TemplateFreeNeuralNetScorer.
Args:
**kwargs: Unused.
"""
from makeit.synthetic.evaluation.rexgen_direct.predict import TFFP
self.model = TFFP()
def get_wln_resutls(self,
reactants_smiles,
contexts=[(20, '', '', '', '', '')],
**kwargs):
"""Evaluates possible reaction outcomes for given reactants.
Evaluation does not use context, but left as dummy pos var.
Args:
reactants_smiles (list of str??): SMILES string of reactants.
contexts (list, optional): Unused.
(default: {[(20,'','','','','')]})
**kwargs: Additional optional parameters.
Returns:
list: Predicted reaction outcomes.
"""
all_outcomes = []
all_mapped_rct_smiles = []
for (T1, slvt1, rgt1, cat1, t1, y1) in contexts:
this_reactants_smiles = reactants_smiles
if slvt1:
this_reactants_smiles += '.' + slvt1
if rgt1:
this_reactants_smiles += '.' + rgt1
if cat1:
this_reactants_smiles += '.' + cat1
mapped_rct_smiles, outcomes = self.model.predict(
this_reactants_smiles, top_n=1e10, atommap=True)
if not outcomes:
all_outcomes.append([{
'rank': 1,
'outcome': {
'smiles': '',
'template_ids': [],
'num_examples': 0
},
'score': 0,
'prob': 0,
}])
continue
outcomes_to_ret = {}
reactants_smiles_split = this_reactants_smiles.split('.')
for outcome in outcomes:
smiles_list = set(self.remove_am(outcome['smiles']).split('.'))
# Canonicalize
smiles_canonical = set()
for smi in smiles_list:
mol = Chem.MolFromSmiles(smi)
if not mol:
MyLogger.print_and_log(
'Template free evaluator could not reparse product {}'
.format('.'.join(smiles_list)),
template_free_scorer_loc, 1)
continue
smiles_canonical.add(Chem.MolToSmiles(mol))
# Remove unreacted frags
smiles_canonical = smiles_canonical - set(
reactants_smiles_split)
if not smiles_canonical:
continue # no reaction?
smiles = max(
smiles_canonical, key=len
) # NOTE: this is not great...byproducts may be longer
if not smiles:
continue
smiles_w_am = ''
for item in outcome['smiles'].split('.'):
# print(item)
try:
mol = Chem.MolFromSmiles(self.remove_am(item))
# print(Chem.MolToSmiles(mol))
if Chem.MolToSmiles(mol) == smiles:
smiles_w_am = item
break
except Exception as e:
print(e)
pass
if smiles in outcomes_to_ret:
outcomes_to_ret[smiles]['rank'] = min(
outcomes_to_ret[smiles]['rank'], outcome['rank'])
outcomes_to_ret[smiles]['score'] = np.log(
np.exp(outcomes_to_ret[smiles]['score']) +
np.exp(outcome['score']))
outcomes_to_ret[smiles]['prob'] += outcome['prob']
else:
# Append outcome information
outcomes_to_ret[smiles] = {
'rank':
outcome['rank'],
'outcome': {
'smiles': smiles,
'smiles_w_am': smiles_w_am,
'template_ids': [],
'num_examples': 0,
},
'score':
float(outcome['score']),
'prob':
float(outcome['prob']),
'mol_wt':
float(Descriptors.MolWt(Chem.MolFromSmiles(smiles)))
}
# Renormalize and re-rank
outcomes = sorted(outcomes_to_ret.values(),
key=lambda x: x['prob'],
reverse=True)
total_prob = sum([outcome['prob'] for outcome in outcomes])
for i, outcome in enumerate(outcomes):
outcomes[i]['rank'] = i + 1
outcomes[i]['prob'] = outcome['prob'] / total_prob
all_outcomes.append(outcomes)
all_mapped_rct_smiles.append(mapped_rct_smiles)
# Return in lists as if we received a list of contexts
return all_mapped_rct_smiles, all_outcomes
def evaluate(self, rxnsmi):
"""only one product allowed at this point"""
# [rsmi, psmi] = rxnsmi.split('>>')
rsmi, resmi, psmi = rxnsmi.split('>')
rsmi_canonical = self.remove_am(rsmi)
psmi_canonical = self.remove_am(psmi)
rsmi_am, outcomes = self.get_wln_resutls(rsmi_canonical)
psmi_am = ''
for outcome in outcomes[0]:
# print(outcome['outcome']['smiles'])
# print(self.remove_am(outcome['outcome']['smiles']))
if outcome['outcome']['smiles'] == psmi_canonical:
psmi_am = outcome['outcome']['smiles_w_am']
if not psmi_am:
print('Failed to find the atom mapping.')
return '>>'
else:
return rsmi_am[0] + '>' + resmi + '>' + psmi_am
def remove_am(self, smi_am):
mol = Chem.MolFromSmiles(smi_am)
for atom in mol.GetAtoms():
# get rid of clearing atom mapping function, so the output will have atom maps
atom.ClearProp('molAtomMapNumber')
# since wln doesn't have stereochemistry, we remove stereochemistry here
rdmolops.RemoveStereochemistry(mol)
return Chem.MolToSmiles(mol)