基于随机森林（RF）的机器学习模型预测hERG阻断剂活性

从分子相似性评估到使用机器学习技术的定量构效关系分析各种建模方法已应用于不同大小和组成的数据集（阻断剂和非阻滞剂的数量）。本研究中使用从公共生物活性数据开发用于预测hERG阻断剂的稳健分类器。随机森林被用来开发使用不同分子描述符，活性阈值和训练集合成的预测模型。与先前提取数据集的研究报告相比，该模型在外部验证中表现出优异的性能。

代码示例

代码语言：javascript复制

#导入依赖库import pandas as pdimport numpy as npimport warnings; warnings.simplefilter('ignore')
from rdkit import Chem, DataStructsfrom rdkit.Chem.Draw import IPythonConsolefrom rdkit.Chem import PandasToolsfrom rdkit.Chem import AllChem, Draw
from sklearn.ensemble import RandomForestClassifier#from sklearn.model_selection import StratifiedKFoldfrom imblearn.under_sampling import RandomUnderSamplerfrom sklearn.metrics import recall_score, roc_auc_scorefrom sklearn.model_selection import KFold, StratifiedKFold,StratifiedShuffleSplitfrom sklearn.model_selection import train_test_splitfrom matplotlib import cm
import mathimport pickleimport os

定义函数

代码语言：javascript复制

class FP:    """    A fingerprint class that inserts molecular fingerprints into pandas data frame    """    def __init__(self, fp):        self.fp = fp    def __str__(self):        return "%d bit FP" % len(self.fp)    def __len__(self):        return len(self.fp)
def get_morgan_fp(mol):    """    Returns the RDKit Morgan fingerprint for a molecule    """    info = {}    arr = np.zeros((1,))    fp = AllChem.GetMorganFingerprintAsBitVect(mol, 2, nBits=1024, useFeatures=False, bitInfo=info)    DataStructs.ConvertToNumpyArray(fp, arr)    arr = np.array([len(info[x]) if x in info else 0 for x in range(1024)])
    return FP(arr)

数据预处理

代码语言：javascript复制

df = pd.read_csv("chembl_training_T3.csv", index_col=0)
PandasTools.AddMoleculeColumnToFrame(df, smilesCol='can_smiles')
df = df[~df.ROMol.isnull()]
df['fp'] = df.apply(lambda x: get_morgan_fp(x['ROMol']), axis=1)

代码语言：javascript复制

df.head()   #查看数据

定义X 、Y（指纹数据集）

代码语言：javascript复制

X = np.array([x.fp for x in df.fp])X.shape
y = np.array(df.ac)y.shape

交叉验证（Cross Validation）

代码语言：javascript复制

# Initialize performance measuressens     = np.array([])spec     = np.array([])auc      = np.array([])
# 10-fold cross-validation splitkfolds = StratifiedKFold(n_splits=10, shuffle=True, random_state=0)kfolds.get_n_splits(X, y)print(kfolds)

代码语言：javascript复制

for train, test in kfolds.split(X, y):    # Split data to training and test set    X_train, X_test, y_train, y_test = X[train], X[test], y[train], y[test]    
    # Training a random forest classifier    rf_clf = RandomForestClassifier(n_estimators=100, criterion='gini', n_jobs=1)    rf_clf.fit(X_train, y_train)
    # Predicting the test set    y_pred       = rf_clf.predict(X_test)    y_pred_proba = rf_clf.predict_proba(X_test).T[1]
    # Append performance measures    auc  = np.append(auc, roc_auc_score(y_test, y_pred_proba))    sens = np.append(sens, recall_score(y_test, y_pred, pos_label=1))    spec = np.append(spec, recall_score(y_test, y_pred, pos_label=0))

# 10-fold cross-validation performanceprint('AUC:ttt%.2f  /- %.2f' % (auc.mean(), auc.std()))print('Sensitivity:tt%.2f  /- %.2f' % (sens.mean(), sens.std()))print('Specificity:tt%.2f  /- %.2f' % (spec.mean(), spec.std()))

代码语言：javascript复制

AUC:			0.95  /- 0.01
Sensitivity:		0.84  /- 0.03
Specificity:		0.91  /- 0.03

测试预测模型（单个分子）

代码语言：javascript复制

mySMILES ='Fc1ccc(cc1)n3c2ccc(Cl)cc2c(c3)C5CCN(CCN4C(=O)NCC4)CC5'from rdkit import RDConfigmySMILESinput = pd.DataFrame(columns=['ID','my_smiles']) mySMILESinput = mySMILESinput.append({'ID':123, 'my_smiles':mySMILES}, ignore_index=True)   PandasTools.AddMoleculeColumnToFrame(mySMILESinput,'my_smiles','ROMol')

代码语言：javascript复制

mySMILESinput['fp'] = mySMILESinput.apply(lambda x: get_morgan_fp(x['ROMol']), axis=1)mySMILESinput

代码语言：javascript复制

resQuery = np.array([x.fp for x in mySMILESinput.fp])y_pred = rf_clf.predict(resQuery)y_pred

代码语言：javascript复制

print(rf_clf.predict(resQuery))print(rf_clf.predict_proba(resQuery))

代码语言：javascript复制

[1]
[[0.04 0.96]]

---

参考：https://github.com/

validation

0 人点赞