RDKit | 基于化合物结构式图像估算分子式

1

简介

当通过深度学习输入有机物质中结构式的二维图像时，需要解决寻找分子式的问题。这是一个回归问题，需要计算结构式图像中包含的碳、氢、氧和氮等原子数。

2

环境

系统 ：Win10

工具：RDKit、OpenCV、Keras、TensorFlow

3

实验步骤

• 训练数据300,000种化合物的SMILES字符串（足够的训练数据）。

• 输入结构的二维图像可以提供足够的信息来理解分子结构。

使用RDKit将SMILES字符串转换为结构式图像，并进行学习以计算图像中的原子数。

4

数据预处理

提取获得类似如下数据

5

示例代码

导入库

import keras
from keras.layers import Input, Concatenate, GlobalAveragePooling2D
from keras.models import Model
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.callbacks import LearningRateScheduler, ModelCheckpoint
from keras.preprocessing.image import img_to_array, array_to_img
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
import os.path
import re
import tensorflow as tf
from keras import backend as K
import cv2
from rdkit import Chem
from rdkit.Chem import Draw, rdDepictor, rdMolDescriptors
import warnings
warnings.filterwarnings("ignore")

载入数据

#Load SMILES
df = pd.read_csv('data.csv')
SMILES = df['CAN_SMILES'].values

SMILES转numpy

# convert to array
SMILES = np.asarray(SMILES)
SMILES_train, SMILES_test = train_test_split(SMILES, test_size=0.30, random_state=110)
print(SMILES_train.shape, SMILES_test.shape)

(210000,) (90000,) 

定义DataGenerator数据生成函数

#Release memory for read data
del df, SMILES
class DataGenerator():
    def __init__(self, X_input):
        self.X_input = X_input
        self.atom = ['C', 'H', 'N', 'O', 'S', 'P', 'Si', 'Na', 'F', 'Cl', 'Br', 'I']
        self.reset()
    def reset(self):
        self.X = []
        self.Y = []
    def cal_img(self, mol):
        # Generate structural formula image
        img = Chem.Draw.MolToImage(mol, size=(300, 300))
        return img_to_array(img)[:,:,:3]
    def chem_count(self, input):
        atom = ['C', 'H', 'N', 'O', 'S', 'P', 'Si', 'Na', 'F', 'Cl', 'Br', 'I']
        count = {'C': 0, 'H': 0 ,'N': 0, 'O': 0, 'S': 0, 'P': 0, 'Si': 0, 'Na': 0, 'F': 0, 'Cl': 0, 'Br': 0, 'I': 0}
        input = input + 'Z' # Add English capital letters at the end
        for i in range(len(atom)):
            if re.search(atom[i] + '\d{1,3}[A-Z]', input):          #Atom + number (1 to 3 digits) + uppercase English
                iterator = re.finditer(atom[i] + '\d{1,3}[A-Z]', input)
                for match in iterator:
                     count[atom[i]] = int(match.group()[len(atom[i]):-1])
            elif re.search(atom[i]  + '[A-Z]',input):              #When containing atoms + uppercase English
                 count[atom[i]] = 1
        return count
    def make_structural_image(self, batch_size=10):
        while True:
            for i in range(len(self.X_input)):
                # SConvert SMILES to mol
                mol = Chem.MolFromSmiles(self.X_input[i], sanitize=True)
                if not mol is None: #  mol file is None
                    # Calculate 2D coordinates
                    rdDepictor.Compute2DCoords(mol)
                    # CalcMolWt
                    MW = Chem.rdMolDescriptors._CalcMolWt(mol)
                    # Calculate molecular formula from mol format
                    chem_Formula = Chem.rdMolDescriptors.CalcMolFormula(mol)
                    # Count atoms from molecular formula
                    count = self.chem_count(chem_Formula)
                    # Add X, Y
                    if MW < max_MW and count['C'] > 1.0: # A compound containing at least two carbon atoms and having a molar mass of 400 or less.
                        (self.X).append(self.cal_img(mol))
                        (self.Y).append([count[self.atom[i]] for i in range(len(self.atom))])
                    # Returns the value of X, Y
                    if len(self.X) == batch_size:
                        inputs = np.asarray(self.X)
                        outputs = np.asarray(self.Y)
                        self.reset()
                        yield inputs, outputs

建立CNN

使用RDKit读取SMILES数据并将其转换为（300 ，300 ，3 ）大小的图像。同样，输出分子式['C'，'H'，'N'，'O'，'S'，'P'，'Si'，'Na'，'F'，'Cl'，'Br' ，'I']是12种原子的数量。

from keras.applications.xception import Xception
inputs = Input(shape=(300,300,3)) # data_size (300,300,3)
x = Xception(weights=None,input_shape=(300,300,3), include_top=False, pooling='avg')(inputs)
y = Dense(12, activation='linear')(x)
model = Model(inputs, y)

训练模型并绘制训练曲线

# Learning
history = model.fit_generator(
    generator=train_datagen.make_structural_image(batch_size=batchsize),
    steps_per_epoch=int(len(SMILES_train) / batchsize / 100),
    epochs=epochs,
    verbose=1,
    validation_data=test_datagen.make_structural_image(batch_size=batchsize),
    validation_steps=100,
    callbacks=callbacks)

#Plot training graph
plt.figure()
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.xlabel('epoch')
plt.ylabel('loss')
plt.legend(['loss', 'val_loss'], loc='upper right')
plt.ylim([0,0.5])
plt.show()

预测测试结果输出

#Output 50 test results
atom = ['C', 'H', 'N', 'O', 'S', 'P', 'Si', 'Na', 'F', 'Cl', 'Br', 'I']
for i in range(50):
    img = array_to_img(X_test[i])
    real = ''
    pred = ''
    for j in range(len(atom)):
        if round(Y_test[i,j]) > 0:
            real += atom[j] + str(round(Y_test[i,j]))
        if round(Y_predict_test[i,j],1) > 0.5:
            pred += atom[j] + str(round(Y_predict_test[i,j],1))
    img.save('test_result%04d_%s_%s.png' % (i, real, pred))

参考资料

1. http://www.rdkit.org/docs/GettingStartedInPython.html

2.https://blog.csdn.net/u012325865/article/details/82318084utm_source=blogxgwz3

3. https://blog.csdn.net/u012325865/article/details/103044562

4. https://blog.csdn.net/qq_35082030/article/details/79144374

5. https://blog.csdn.net/qq_19332527/article/details/79829087

作者&编辑

王建民

DrugAI

长按识别二维码关注我们获取最新消息！

本文为DrugAI原创编译整理，如需转载，请在公众号后台留言。