def get_score(src, trg, task='logp04'):
if task == 'logp04':
logp_improvement = penalized_logp(trg) - penalized_logp(src)
# return logp_improvement * (0.6+min(sim, 0.4))
return logp_improvement
elif task == 'qed':
return qed(trg)
elif task == 'drd2':
try:
val = drd2(trg)
except:
print(
"***************\n***************\n***************\n***************\n"
)
print(
"***************\n***************\n***************\n***************\n"
)
print(trg)
print(
"***************\n***************\n***************\n***************\n"
)
print(
"***************\n***************\n***************\n***************\n"
)
val = 0
return val
else:
assert 'wrong task: %s' % task
def get_scores(src, trg):
sim = similarity(src, trg)
# if sim < 0.4 or sim >= 1.0:
if sim >= 1.0:
sim = 0
logp_improvement = penalized_logp(trg) - penalized_logp(src)
qed_score = qed(trg)
drd2_score = drd2(trg)
return sim, logp_improvement, qed_score, drd2_score
def get_score(src, trg, task='logp04'):
sim = similarity(src, trg)
if sim >= 1.0:
return 0
if task == 'logp04':
logp_improvement = penalized_logp(trg) - penalized_logp(src)
# return logp_improvement * (0.6+min(sim, 0.4))
if sim < 0.4 and sim >= 0:
return 0
else:
return logp_improvement
elif task == 'qed':
if sim < 0.4:
return 0
else:
return qed(trg)
elif task == 'drd2':
if sim < 0.4:
return 0
else:
try:
val = drd2(trg)
except:
print(
"***************\n***************\n***************\n***************\n"
)
print(
"***************\n***************\n***************\n***************\n"
)
print(trg)
print(
"***************\n***************\n***************\n***************\n"
)
print(
"***************\n***************\n***************\n***************\n"
)
val = 0
return val
else:
assert 'wrong task: %s' % task
def evaluate(test_smiles_list, model, moltokenizer, log_diff_list,
qed_diff_list, drd2_diff_list):
smiles_list = []
ids_list = []
property_x_list = []
property_desired_list = []
n_grid = 3 * 3 * 1
for idx, smiles in enumerate(test_smiles_list):
ids = moltokenizer.encode(smiles)
logp_val = penalized_logp(smiles)
qed_val = qed(smiles)
drd2_val = drd2(smiles)
property_x = np.expand_dims(np.array([logp_val, qed_val, drd2_val]),
axis=0)
for log_diff in log_diff_list:
logp_improvement = logp_val + log_diff
for qed_diff in qed_diff_list:
qed_desired = qed_val + qed_diff
for drd2_diff in drd2_diff_list:
drd2_desired = drd2_val + drd2_diff
property_desired = np.expand_dims(np.array(
[logp_improvement, qed_desired, drd2_desired]),
axis=0)
smiles_list.append(smiles)
ids_list.append(ids)
property_x_list.append(property_x)
property_desired_list.append(property_desired)
x = np.array(
tf.keras.preprocessing.sequence.pad_sequences(ids_list,
dtype="int64",
padding="post"))
px = np.concatenate(property_x_list, axis=0)
py = np.concatenate(property_desired_list, axis=0)
print(px[2, :])
print(py[2, :])
print(px[10, :])
print(py[10, :])
outputs, _ = model.predict([x, px, py], batch_size=args.eb, verbose=1)
if args.p == 1:
valid_smiles_list = []
score_plogp_all = []
score_qed_all = []
score_drd2_all = []
for idx, output in tqdm.tqdm(enumerate(outputs)):
smiles_x = smiles_list[idx]
smiles_y = moltokenizer.decode(output)
sim = similarity(smiles_x, smiles_y)
if sim < 0.4:
score_plogp_all.append(0)
score_qed_all.append(0)
score_drd2_all.append(0)
valid_smiles_list.append('')
else:
score_plogp = get_score(smiles_x, smiles_y, task='logp04')
score_qed = get_score(smiles_x, smiles_y, task='qed')
score_drd2 = get_score(smiles_x, smiles_y, task='drd2')
score_plogp_all.append(score_plogp)
score_qed_all.append(score_qed)
score_drd2_all.append(score_drd2)
valid_smiles_list.append(smiles_y)
else:
with Timer("score multi calculation..."):
n = len(outputs)
n_proc = args.p
batch = math.ceil(n / (n_proc))
with Pool(processes=n_proc) as pool:
r = pool.map_async(
worker_wrapper,
zip(repeat(outputs), repeat(smiles_list), repeat(n),
range(1, n_proc + 1), range(0, n, batch),
range(batch, batch * n_proc + 1, batch)))
r.wait()
score_all = []
valid_smiles_list = []
for partial_score, valid_smiles_p_list in r.get():
score_all += partial_score
valid_smiles_list += valid_smiles_p_list
score_all = np.array(score_all)
score_plogp_all = score_all[:, 0]
score_qed_all = score_all[:, 1]
score_drd2_all = score_all[:, 2]
return score_plogp_all, score_qed_all, score_drd2_all, valid_smiles_list
)
moltokenizer = Moltokenizer(params["vocab_file"])
epoch_num = 270 # drd2 test 0.61, V0011
# epoch_num = 210
tst_model = get_model(params, epoch_num)
selected_smiles = 'COC1=CC=C(C=C1)C(=O)N1CCCC1=O'
log_diff_list = [0.0]
qed_diff_list = [-0.1, -0.05, 0.0, 0.05, 0.1]
drd2_diff_list = np.array(range(6, 10)) / 10.0
org_plp = penalized_logp(selected_smiles)
org_qed = qed(selected_smiles)
org_drd2 = drd2(selected_smiles)
score_plogp_all, score_qed_all, score_drd2_all, valid_smiles_list = \
evaluate([selected_smiles], tst_model, moltokenizer, log_diff_list, qed_diff_list, drd2_diff_list)
new_plp = score_plogp_all[np.argmax(score_drd2_all)]
new_qed = score_qed_all[np.argmax(score_drd2_all)]
new_drd2 = score_drd2_all[np.argmax(score_drd2_all)]
new_smiles = valid_smiles_list[np.argmax(score_drd2_all)]
print(org_plp, org_qed, org_drd2)
print(new_plp, new_qed, new_drd2)
print("improvemend:", new_drd2 - org_drd2)
print("diff:", (new_plp - org_plp)**2 + (new_qed - org_qed)**2)
print(new_smiles)