|
import numpy as np
|
|
import pandas as pd
|
|
import pickle
|
|
import hydra
|
|
|
|
import warnings
|
|
warnings.filterwarnings('ignore')
|
|
|
|
import rdkit.Chem as Chem
|
|
from rdkit import RDLogger
|
|
RDLogger.DisableLog('rdApp.*')
|
|
from rdkit.Chem import AllChem, QED
|
|
|
|
def getReward(name,
|
|
receptor_path=None,
|
|
pdbqt_path=None,
|
|
VinaGPU_path=None,
|
|
VinaGPU_config=None):
|
|
if name == "QED":
|
|
return QEDReward()
|
|
elif name == 'DRD2':
|
|
with open(hydra.utils.get_original_cwd() + '/Model/QSAR/drd2_qsar_optimized.pkl', mode='rb') as f:
|
|
qsar_model = pickle.load(f)
|
|
return QSAR_Reward(qsar_model)
|
|
elif name == 'AKT1':
|
|
with open(hydra.utils.get_original_cwd() + '/Model/QSAR/akt1_qsar_optimized.pkl', mode='rb') as f:
|
|
qsar_model = pickle.load(f)
|
|
return QSAR_Reward(qsar_model)
|
|
|
|
|
|
class Reward:
|
|
def __init__(self):
|
|
self.vmin = -100
|
|
self.max_r = -10000
|
|
return
|
|
|
|
def reward(self):
|
|
raise NotImplementedError()
|
|
|
|
class QSAR_Reward(Reward):
|
|
def __init__(self, qsar_model, *args, **kwargs):
|
|
super().__init__(*args, **kwargs)
|
|
self.qsar_model = qsar_model
|
|
|
|
def reward(self, score_que:list = None):
|
|
max_smi = None
|
|
scores = []
|
|
mols = [Chem.MolFromSmiles(smi) for smi in score_que]
|
|
ecfps = []
|
|
None_indices = []
|
|
for i, mol in enumerate(mols):
|
|
if mol is not None:
|
|
ecfps.append(AllChem.GetMorganFingerprintAsBitVect(mol, 3, nBits=2048))
|
|
else:
|
|
None_indices.append(i)
|
|
ecfps.append([0]*2048)
|
|
if len(ecfps) == 0:
|
|
return [], None, None
|
|
ecfp6_array = np.array(ecfps)
|
|
X = pd.DataFrame(ecfp6_array, columns=[f'bit_{i}' for i in range(2048)])
|
|
y_pred = self.qsar_model.predict_proba(X)[:, 1]
|
|
for None_idx in None_indices:
|
|
y_pred[None_idx] = np.nan
|
|
max_score = np.nanmax(y_pred)
|
|
for smi, score in zip(score_que, y_pred):
|
|
if score == np.nan:
|
|
pass
|
|
elif score == max_score:
|
|
max_smi = smi
|
|
scores.append((smi, score))
|
|
return scores, max_smi, max_score
|
|
|
|
def reward_remove_nan(self, score_que:list = None):
|
|
max_smi = None
|
|
scores = []
|
|
|
|
valid_smiles = []
|
|
mols = []
|
|
for smi in score_que:
|
|
mol = Chem.MolFromSmiles(smi)
|
|
if mol is not None:
|
|
valid_smiles.append(smi)
|
|
mols.append(mol)
|
|
ecfps = []
|
|
for i, mol in enumerate(mols):
|
|
if mol is not None:
|
|
ecfps.append(AllChem.GetMorganFingerprintAsBitVect(mol, 3, nBits=2048))
|
|
if len(ecfps) == 0:
|
|
return [], None, None
|
|
ecfp6_array = np.array(ecfps)
|
|
X = pd.DataFrame(ecfp6_array, columns=[f'bit_{i}' for i in range(2048)])
|
|
y_pred = self.qsar_model.predict_proba(X)[:, 1]
|
|
max_score = np.nanmax(y_pred)
|
|
for smi, score in zip(valid_smiles, y_pred):
|
|
if score == max_score:
|
|
max_smi = smi
|
|
scores.append((smi, score))
|
|
return scores, max_smi, max_score
|
|
|
|
class QEDReward(Reward):
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__(*args, **kwargs)
|
|
self.vmin = 0
|
|
|
|
def reward(self, smi):
|
|
mol = Chem.MolFromSmiles(smi)
|
|
try:
|
|
score = QED.qed(mol)
|
|
except:
|
|
score = None
|
|
|
|
return score |
