def retrain(self, n, keep_top_n, smiles_and_scores):
print("writing dataset...")
name = 'molexit-%d' % n
dataset = '../models/molexit/%s.txt' % name
dataset_scores = []
with open(dataset, 'w') as f:
for smi, score in list(
reversed(sorted(smiles_and_scores,
key=lambda p: p[1])))[:keep_top_n]:
dsmi = self.converter.encode(
pybel.readstring("smi", smi.strip()).write("can").strip())
tok = DeepSMILESTokenizer(dsmi)
tokens = tok.get_tokens()
f.write(' '.join([t.value for t in tokens]))
f.write("\n")
dataset_scores.append(score)
print('dataset: size: %s, mean score: %s, max score: %s' %
(len(dataset_scores), np.mean(dataset_scores),
np.max(dataset_scores)))
print('training new LM...')
self.lm_trainer.train(10, dataset, '../models/molexit', name)
vocab = get_arpa_vocab('../models/molexit/%s.arpa' % name)
self.lm = KenLMDeepSMILESLanguageModel(
'../models/molexit/%s.klm' % name, vocab)
def __init__(self):
self.vocab = get_arpa_vocab(
'../resources/chembl_25_deepsmiles_klm_10gram_200503.arpa')
self.lm = KenLMDeepSMILESLanguageModel(
'../resources/chembl_25_deepsmiles_klm_10gram_200503.klm',
self.vocab)
self.best_smiles = None
self.best_score = -1.0
def generate_optimized_molecules(self,
scoring_function,
number_molecules,
starting_population=None):
self.new_model_dir()
vocab = get_arpa_vocab(
'../resources/chembl_25_deepsmiles_klm_10gram_200503.arpa')
self.lm = KenLMDeepSMILESLanguageModel(
'../resources/chembl_25_deepsmiles_klm_10gram_200503.klm', vocab)
print("generating %s samples..." % number_molecules)
smiles_and_scores = []
TIME_PER_ITERATION = self.time_per_iteration_minutes * 60 # in seconds
found = False
for n in range(1, self.num_iterations + 1):
print("iteration %s" % n)
num_valid = 0
start = time.time()
elapsed = time.time() - start
while elapsed < TIME_PER_ITERATION:
try:
generated = self.lm.generate(num_chars=100,
text_seed='<s>')
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
smiles = DeepSMILESLanguageModelUtils.sanitize(decoded)
score = scoring_function.score(smiles)
num_valid += 1
smiles_and_scores.append((smiles, score))
if score == 1.0:
found = True
break
except Exception:
pass
elapsed = time.time() - start
print("num valid: %s" % num_valid)
if found:
break
self.retrain(n, self.keep_top_n, smiles_and_scores)
return [
pair[0] for pair in list(
reversed(sorted(smiles_and_scores, key=lambda p: p[1])))
[:number_molecules]
]
from rdkit.Chem.Crippen import MolLogP
from rdkit.RDLogger import logger
from rdkit import Chem
from chemgrams import get_arpa_vocab, EmptyDeepSMILESLanguageModel, DeepSMILESLanguageModelUtils
from rdkit import rdBase
rdBase.DisableLog('rdApp.error')
rdBase.DisableLog('rdApp.warning')
logger = logger()
vocab = get_arpa_vocab(
'../resources/chemts_250k_deepsmiles_klm_6gram_190414.arpa')
lm = EmptyDeepSMILESLanguageModel(vocab, n=6)
current_best_score = None
current_best_smiles = None
beats_current = lambda score: score < current_best_score
for i in range(1000):
generated = lm.generate(num_chars=25, text_seed="<s>")
try:
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
sanitized = DeepSMILESLanguageModelUtils.sanitize(decoded)
mol = Chem.MolFromSmiles(sanitized)
logp_score = MolLogP(mol)
logger.info(
"KenLMDeepSMILESLanguageModel('../resources/zinc12_fragments_deepsmiles_klm_10gram_200502.klm', vocab)"
)
logger.info("width = 12, max_depth = 35, start_state = ['<s>'], c = 5")
logger.info(
"score: -1.0 if invalid; -1.0 if seen previously; tanimoto distance from abilify if valid"
)
logger.info("LanguageModelMCTSWithPUCTTerminating")
logger.info("TanimotoScorer(abilify, radius=6)")
logger.info("num_iterations = 100")
logger.info("simulations_per_iteration = 50000")
logger.info("keep_top_n = 5000")
logger.info("loading language model...")
vocab = get_arpa_vocab(
'../resources/zinc12_fragments_deepsmiles_klm_10gram_200502.arpa')
lm = KenLMDeepSMILESLanguageModel(
'../resources/zinc12_fragments_deepsmiles_klm_10gram_200502.klm', vocab)
abilify = "Clc4cccc(N3CCN(CCCCOc2ccc1c(NC(=O)CC1)c2)CC3)c4Cl"
scorer = TanimotoScorer(abilify, radius=6)
converter = Converter(rings=True, branches=True)
env = os.environ.copy()
env["PATH"] = "/Users/luis/kenlm/build/bin:" + env["PATH"]
lm_trainer = KenLMTrainer(env)
def log_best(j, all_best, n_valid, lggr):
if j % 1000 == 0:
lggr.info("--iteration: %d--" % j)
from rdkit import rdBase
rdBase.DisableLog('rdApp.error')
rdBase.DisableLog('rdApp.warning')
logger = get_logger('chemgrams.log')
THIS_DIR = os.path.dirname(os.path.abspath(__file__))
logger.info("LM-only")
logger.info(
"KenLMDeepSMILESLanguageModel(n=10, 'chemts_250k_deepsmiles_klm_10gram_200429.klm')"
)
logger.info("num_chars=100, text_seed='<s>'")
vocab = get_arpa_vocab(
'../resources/chemts_250k_deepsmiles_klm_10gram_200429.arpa')
lm = KenLMDeepSMILESLanguageModel(
'../resources/chemts_250k_deepsmiles_klm_10gram_200429.klm', vocab)
all_smiles = set()
num_valid = 0
start = time.time()
for i in range(500000): # about enough to get ~250,000 valid molecules
try:
generated = lm.generate(num_chars=100, text_seed='<s>')
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
sanitized = DeepSMILESLanguageModelUtils.sanitize(decoded)
logger.info(
"KenLMDeepSMILESLanguageModel('../resources/chembl_25_deepsmiles_klm_10gram_200503.klm', vocab)"
)
logger.info("width = 12, max_depth = 50, start_state = ['<s>'], c = 15")
logger.info(
"score: -1.0 if invalid; -1.0 if seen in iteration; tanimoto distance from abilify if valid; rescaling from [0,1] to [-1,1]"
)
logger.info("LanguageModelMCTSWithPUCTTerminating")
logger.info(
"TanimotoScorer(abilify, radius=6); distance only (no SA or cycle scoring)"
)
logger.info("num_iterations = 100")
logger.info("time per iteration = 45 min.")
logger.info("keep_top_n = 20000 unique")
vocab = get_arpa_vocab(
'../resources/chembl_25_deepsmiles_klm_10gram_200503.arpa')
lm = KenLMDeepSMILESLanguageModel(
'../resources/chembl_25_deepsmiles_klm_10gram_200503.klm', vocab)
abilify = "Clc4cccc(N3CCN(CCCCOc2ccc1c(NC(=O)CC1)c2)CC3)c4Cl"
distance_scorer = TanimotoScorer(abilify, radius=6)
converter = Converter(rings=True, branches=True)
env = os.environ.copy()
env["PATH"] = "/Users/luis/kenlm/build/bin:" + env["PATH"]
lm_trainer = KenLMTrainer(env)
def smiles_to_deepsmiles(smiles):
canonical = pybel.readstring("smi", smiles).write("can").strip()
return converter.encode(canonical)
logger.info("KenLMDeepSMILESLanguageModel('../resources/chembl_25_deepsmiles_klm_10gram_200503.klm', vocab)")
logger.info("width = 12, max_depth = 50, start_state = ['<s>'], c = 5")
logger.info("score: -1.0 if invalid; -1.0 if seen previously; TanimotoScorer(abilify, radius=6) if valid; rescaling from [0,1] to [-1,1]")
logger.info("LanguageModelMCTSWithPUCTTerminating")
TIME_LIMIT = 3 * 60 * 60 # three hours in seconds
# TIME_LIMIT = 2*60 # 2 minutes in seconds
LOG_INTERVAL = 1 * 60 * 60 # one hour in seconds
# LOG_INTERVAL = 30.0 # 30 seconds
KEEP_TOP_N = 20000
logger.info("loading language model...")
vocab = get_arpa_vocab('../resources/chembl_25_deepsmiles_klm_10gram_200503.arpa')
lm = KenLMDeepSMILESLanguageModel('../resources/chembl_25_deepsmiles_klm_10gram_200503.klm', vocab)
abilify = "Clc4cccc(N3CCN(CCCCOc2ccc1c(NC(=O)CC1)c2)CC3)c4Cl"
distance_scorer = TanimotoScorer(abilify, radius=6)
num_simulations = 15000000 # much more than 8 hours
width = 12
max_depth = 50
start_state = ["<s>"]
c = 5
all_unique = {}
all_valid = []
num_valid = 0
simulations = 0
best = mcts.get_best_sequence()
generated_text = ''.join(best[0])
logger.info("best generated text: %s" % generated_text)
decoded = DeepSMILESLanguageModelUtils.decode(generated_text,
start='<s>',
end='</s>')
smiles = DeepSMILESLanguageModelUtils.sanitize(decoded)
logger.info("best SMILES: %s, J: %s (%s seconds)" %
(smiles, distance_scorer.score(smiles), str((end - start))))
log_top_best(all_smiles, 5, logger)
logger.info("writing dataset...")
name = 'molexit-%d' % n
dataset = '../models/molexit/%s.txt' % name
with open(dataset, 'w') as f:
for smi in list(
reversed(sorted(all_smiles.items(),
key=lambda kv: kv[1][0])))[:keep_top_n]:
dsmi = smiles_to_deepsmiles(smi[0].strip())
tok = DeepSMILESTokenizer(dsmi)
tokens = tok.get_tokens()
f.write(' '.join([t.value for t in tokens]))
f.write("\n")
logger.info('training new LM...')
lm_trainer.train(6, dataset, '../models/molexit', name)
vocab = get_arpa_vocab('../models/molexit/%s.arpa' % name)
lm = KenLMDeepSMILESLanguageModel('../models/molexit/%s.klm' % name, vocab)
from chemgrams import get_arpa_vocab, KenLMDeepSMILESLanguageModel, DeepSMILESLanguageModelUtils, \
LanguageModelMCTSWithUCB1
from rdkit.RDLogger import logger
from rdkit import rdBase, Chem
rdBase.DisableLog('rdApp.error')
rdBase.DisableLog('rdApp.warning')
logger = logger()
if __name__ == '__main__':
logger.info("loading language model...")
vocab = get_arpa_vocab(
'../models/chembl_25_deepsmiles_klm_6gram_190413.arpa')
lm = KenLMDeepSMILESLanguageModel(
'../models/chembl_25_deepsmiles_klm_6gram_190413.klm', vocab)
num_simulations = 1000
width = 3
text_length = 25
start_state = ["<s>"]
def eval_function(text):
generated = ''.join(text)
try:
decoded = DeepSMILESLanguageModelUtils.decode(generated,
start='<s>',
end='</s>')
DeepSMILESLanguageModelUtils.sanitize(decoded)
from chemgrams import get_arpa_vocab, KenLMSELFIESLanguageModel
import selfies as sf
from openbabel import pybel
if __name__ == "__main__":
# 6.655716427238491
vocab = get_arpa_vocab(
'../resources/chemts_250k_selfies_klm_10gram_210908.arpa')
lm = KenLMSELFIESLanguageModel(
'../resources/chemts_250k_selfies_klm_10gram_210908.klm', vocab)
def smiles_to_selfies(smiles):
canonical = pybel.readstring("smi", smiles).write("can").strip()
return sf.encoder(canonical)
with open("../resources/zinc12_enaminebb_smiles_corpus.txt") as f:
all_smiles = [s.strip() for s in f.readlines()]
# the sum of log10 probs of each sentence in the corpus
sum_log_prob = 0.0
# the total number of "words" (i.e. tokens) in the corpus
M = 0
for smiles in all_smiles:
s = smiles_to_selfies(smiles.strip())
tokens = list(sf.split_selfies(s))
M += len(tokens)
sum_log_prob += lm.log_prob(' '.join(tokens))
logger.info(
"KenLMSELFIESLanguageModel('../resources/zinc12_fragments_selfies_klm_10gram_210908.klm', vocab)"
)
logger.info("width = 12, max_depth = 35, start_state = ['<s>'], c = 5")
logger.info(
"score: -1.0 if invalid; -1.0 if seen previously; tanimoto distance from abilify if valid"
)
logger.info("LanguageModelMCTSWithPUCTTerminating")
logger.info("TanimotoScorer(abilify, radius=6)")
logger.info("num_iterations = 100")
logger.info("simulations_per_iteration = 50000")
logger.info("keep_top_n = 1500")
logger.info("loading language model...")
vocab = get_arpa_vocab(
'../resources/zinc12_fragments_selfies_klm_10gram_210908.arpa')
lm = KenLMSELFIESLanguageModel(
'../resources/zinc12_fragments_selfies_klm_10gram_210908.klm', vocab)
abilify = "Clc4cccc(N3CCN(CCCCOc2ccc1c(NC(=O)CC1)c2)CC3)c4Cl"
scorer = TanimotoScorer(abilify, radius=6)
converter = Converter(rings=True, branches=True)
env = os.environ.copy()
env["PATH"] = "/Users/luis/kenlm/build/bin:" + env["PATH"]
lm_trainer = KenLMTrainer(env)
def log_best(j, all_best, n_valid, lggr):
if j % 1000 == 0:
lggr.info("--iteration: %d--" % j)