def eval_function(text):
global i, num_valid, all_smiles
i += 1
generated = ''.join(text)
try:
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
smiles = DeepSMILESLanguageModelUtils.sanitize(decoded)
except Exception:
return -1.0
num_valid += 1
if smiles in all_smiles:
return -1.0
else:
# the score in this case is simply 1.0, since the molecule is valid and hasn't been generated yet;
# but it could be anything, such as whether the generated sequence contains sulfur, etc.
score = 1.0
all_smiles.add(smiles)
# As in "Molecular de-novo design through deep reinforcement learning", by Olivecrona et al., we are adding
# the prior's log probability of the generated sequence to the score.
prior_log_prob = prior.log_prob(
DeepSMILESLanguageModelUtils.extract_sentence(text,
join_on=' ',
start='<s>',
end='</s>'))
tot_score = prior_log_prob + sigma * score
# rescale the score
# in practice, the log probs are rarely less than -45; so the min tot_score can be: -45 + (sigma*-1.0)
rescale_min = -45 - sigma
if tot_score < rescale_min:
logger.info("WARNING: total score lower than %s" % rescale_min)
# because probabilities are in the range [0,1], the max log prob is log(1) i.e. 0
# so the max tot_score can be: 0 + sigma*1.0
rescale_max = sigma
# scaling x into [a,b]: (b-a)*((x - min(x))/(max(x) - min(x))+a
ret_score = (1 - (-1)) * ((tot_score - rescale_min) /
(rescale_max - rescale_min)) + (-1)
return ret_score
def eval_function(text):
global simulations, all_unique, elapsed, current_best_score, current_best_smiles, beats_current
if elapsed >= time_limit or len(all_valid) == max_gen:
raise StopTreeSearch()
simulations += 1
generated = ''.join(text)
try:
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
smiles = DeepSMILESLanguageModelUtils.sanitize(decoded)
mol = Chem.MolFromSmiles(smiles)
if mol is None: raise Exception
except Exception:
elapsed = time.time() - start
return -1.0
score = distance_scorer.score_mol(mol)
if current_best_score is None or beats_current(score):
current_best_score = score
current_best_smiles = smiles
if score == 1.0:
logger.info("FOUND!")
###
# As in "Molecular de-novo design through deep reinforcement learning", by Olivecrona et al., we are adding
# the prior's log probability of the generated sequence to the score.
prior_log_prob = prior.log_prob(
DeepSMILESLanguageModelUtils.extract_sentence(text,
join_on=' ',
start='<s>',
end='</s>'))
rescaled_distance_score = (score * 2) + (-1)
# rescale the prior log prob
# in practice, the log probs are rarely less than -45
rescale_min = -45
if prior_log_prob < rescale_min:
logger.info("WARNING: prior log prob lower than %s" % rescale_min)
# because probabilities are in the range [0,1], the max log prob is log(1) i.e. 0
rescale_max = 0.0
# scaling x into [a,b]: (b-a)*((x - min(x))/(max(x) - min(x))+a
rescaled_log_prob = (1 - (-1)) * ((prior_log_prob - rescale_min) /
(rescale_max - rescale_min)) + (-1)
ret_score = (
1 - sigma) * rescaled_log_prob + sigma * rescaled_distance_score
ret_score = -1.0 if smiles in seen else ret_score
###
all_unique[smiles] = (score, generated)
all_valid.append((smiles, score))
seen.add(smiles)
elapsed = time.time() - start
return ret_score
def eval_function(text):
global simulations, num_valid, all_unique, elapsed, current_best_score, current_best_smiles, beats_current
if elapsed >= TIME_PER_ITERATION:
raise StopTreeSearch()
simulations += 1
generated = ''.join(text)
try:
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
smiles = DeepSMILESLanguageModelUtils.sanitize(decoded)
mol = Chem.MolFromSmiles(smiles)
if mol is None: raise Exception
except Exception:
elapsed = time.time() - start
return -1.0
num_valid += 1
distance_score = distance_scorer.score_mol(mol)
if distance_score == 1.0:
logger.info("FOUND!")
# As in "Molecular de-novo design through deep reinforcement learning", by Olivecrona et al., we are adding
# the prior's log probability of the generated sequence to the score.
prior_log_prob = prior.log_prob(
DeepSMILESLanguageModelUtils.extract_sentence(text,
join_on=' ',
start='<s>',
end='</s>'))
# tot_score = prior_log_prob + sigma * ((distance_score * 2) + (-1)) # rescale the distance score from [0,1] to [-1,1]
tot_score = prior_log_prob + sigma * distance_score
# rescale the score
# in practice, the log probs are rarely less than -45; so the min tot_score can be: -45 + (sigma*-1.0)
rescale_min = -45 - sigma
if tot_score < rescale_min:
logger.info("WARNING: total score lower than %s" % rescale_min)
# because probabilities are in the range [0,1], the max log prob is log(1) i.e. 0
# so the max tot_score can be: 0 + sigma*1.0
rescale_max = sigma
# scaling x into [a,b]: (b-a)*((x - min(x))/(max(x) - min(x))+a
ret_score = (1 - (-1)) * ((tot_score - rescale_min) /
(rescale_max - rescale_min)) + (-1)
ret_score = -1.0 if smiles in seen else ret_score
if current_best_score is None or beats_current(distance_score):
current_best_score = distance_score
current_best_smiles = smiles
all_unique[smiles] = (distance_score, generated)
all_valid.append((smiles, distance_score))
seen.add(smiles)
elapsed = time.time() - start
return ret_score
