datacastle 比赛：AI战疫·小分子成药属性预测大赛

比赛入口

任务

根据小分子的属性特征，预测小分子在人体内清除率指标（即数据中的Label字段）。

 

训练数据

>> df.head
<bound method NDFrame.head of         ID  Molecule_max_phase  Molecular weight  RO5_violations     AlogP  \
0     1003                   0          0.206754             0.0  0.608040   
1     1819                   0          0.130056             0.0  0.591848   
2     6090                   0          0.162482             0.0  0.592965   
3     3916                   0          0.112266             0.0  0.652150   
4     8480                   0          0.161722             0.0  0.741485   
...    ...                 ...               ...             ...       ...   
6919  6612                   4          0.101742             0.0  0.621999   
6920  1316                   0          0.227768             1.0  0.790620   
6921    54                   0          0.331437             2.0  0.811837   
6922  3577                   0          0.250387             1.0  0.753769   
6923  5955                   4          0.158947             0.0  0.613065   

                                               Features      Label  
0     [0, 0, 0, 0, 0, 0, 2, 0, 10, 0, 0, 3, 2, 0, 0,...   3.190476  
1     [0, 0, 0, 0, 0, 0, 0, 0, 5, 0, 0, 5, 1, 0, 0, ...   9.740969  
2     [0, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 7, 0, 0, 0, ...  10.545341  
3     [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 9, 0, 0, 0, ...   3.206803  

特征有：

字段名	类型	说明
ID	整型	样本编号
Molecule_max_phase	整型	分子的最长位相
Molecular weight	浮点型	分子量
RO5_violations	整型	违反新药5规则（RO5）的数量
AlogP	浮点型	由ACD软件计算化合物的脂分配系数（该数据来自ChemBL）
Features	向量	小分子的矢量化表示
Label	枚举/浮点型	单位时间内单位机体能将多少容积体液中的药物清除

这里也看不懂，就用线性回归处理下吧

import pandas as pd
df = pd.read_csv('train.csv')
df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 6924 entries, 0 to 6923
Data columns (total 7 columns):
 #   Column              Non-Null Count  Dtype  
---  ------              --------------  -----  
 0   ID                  6924 non-null   int64  
 1   Molecule_max_phase  6924 non-null   int64  
 2   Molecular weight    6924 non-null   float64
 3   RO5_violations      6853 non-null   float64
 4   AlogP               6853 non-null   float64
 5   Features            6924 non-null   object 
 6   Label               6924 non-null   float64
dtypes: float64(4), int64(2), object(1)
memory usage: 378.8+ KB

 

把features 里的特征拿出来和外面4个特征放在一起

import numpy as np
dffeature = df.Features.values
dffeature = np.array(dffeature)
newfeature = []
for line in dffeature:
    line = line[1:-1]
    arra = line.split(',')
    arra = np.array(arra).astype(np.float64)
    newfeature.append(arra)
newfeature = np.array(newfeature)
newfeatureall = np.concatenate((np.array(df.Molecule_max_phase).reshape(df.Molecule_max_phase.shape[0],1),np.array(df['Molecular weight']).reshape(df.Molecule_max_phase.shape[0],1)),axis=1)
newfeatureall = np.concatenate((newfeatureall,np.array(df['RO5_violations']).reshape(df.Molecule_max_phase.shape[0],1)),axis=1)
newfeatureall = np.concatenate((newfeatureall,np.array(df['AlogP']).reshape(df.Molecule_max_phase.shape[0],1)),axis=1)
newfeatureall = np.concatenate((newfeatureall,newfeature),axis=1)
newfeatureall.shape

 

过滤掉基本不变（最小值等于最大值）的特征

selectcolumn = (newfeatureall.max(axis=0) != newfeatureall.min(axis=0))
selectcolumn = np.array(selectcolumn)
newfeatureselect = newfeatureall[:,selectcolumn]

 

特征标准化

from sklearn.preprocessing import StandardScaler
stdScale = StandardScaler().fit(newfeatureselect)
newfeaturenorm = stdScale.transform(newfeatureselect)

 

用岭回归做拟合 

newlabel  = df.Label
newfeaturenorm[np.isnan(newfeaturenorm)]=0
from sklearn.linear_model import Ridge
clf = Ridge(alpha=1.0)
#from sklearn.linear_model import Lasso
#clf = Lasso(alpha=0.01,positive=True)
clf.fit(newfeaturenorm, newlabel)

 

读取测试数据，也做特征过滤和标准化

testdf = pd.read_csv('test_noLabel.csv')
testfeature = np.array(testdf.Features.values)
testfeature.shape
testfeaturearr= []
for line in testfeature:
    line = line[1:-1]
    arra = line.split(',')
    arra = np.array(arra).astype(np.float64)
    testfeaturearr.append(arra)
testfeaturearr = np.array(testfeaturearr)

testfeatureall = np.concatenate((np.array(testdf.Molecule_max_phase).reshape(testdf.Molecule_max_phase.shape[0],1),np.array(testdf['Molecular weight']).reshape(testdf.Molecule_max_phase.shape[0],1)),axis=1)
testfeatureall = np.concatenate((testfeatureall,np.array(testdf['RO5_violations']).reshape(testdf.Molecule_max_phase.shape[0],1)),axis=1)
testfeatureall = np.concatenate((testfeatureall,np.array(testdf['AlogP']).reshape(testdf.Molecule_max_phase.shape[0],1)),axis=1)
testfeatureall = np.concatenate((testfeatureall,testfeaturearr),axis=1)


testfeatureselect = testfeatureall[:,selectcolumn]
testfeaturenorm = stdScale.transform(testfeatureselect)
newfeaturenorm[np.isnan(newfeaturenorm)]=0

 

对测试数据进行预测并保存

result = clf.predict(newfeaturenorm)
result
testid = np.array(testdf.ID.values)
resultfile = open('result.csv','w')
resultfile.write("ID,Label\n")
for index in range(len(testid)):
    result[index] = result[index] if result[index] >= 0 else 0
    resultfile.write(str(testid[index])+","+str(result[index])+"\n")
resultfile.close()

 

评分采用RMSE，结果跑出来6.多，不太理想，再想想哪里可以优化吧。

• 缺失值处理
• 岭回归不同参数例如lumda
• 试下其他回归例如lasso
• 画图来进一步判断哪些特征有用，排除不太相关的特征

 

---

20200313 更新 ，内容：

缺失值采用均值、众数、拉格朗日插值处理；调整岭回归alpha参数，排除ID列，去掉带有nan值的一条记录，rmse来到了3.0多；把可以参数放到代码最前面方便调参

import pandas as pd
import numpy as np
from sklearn.preprocessing import StandardScaler
from scipy.interpolate import lagrange #导入拉格朗日函数

df = pd.read_csv('train_0312.csv')
testdf = pd.read_csv('test_noLabel_0312.csv')
df = df.drop('ID',axis=1)
testid = np.array(testdf.ID.values)
testdf = testdf.drop('ID',axis=1)
newlabel  = df.Label

missdealtype = 3  #缺失值处理方法 1：众数 2：平均值，3：拉格朗日插值
normaltype = 2  #正则化方法，1：标准化  2：归一化
modeltype = 1 #模型类型 ，1：岭回归
alphacoef=2.5
lagelangk = 2

def ployinterp_column(s,n,k=lagelangk): #k=2表示用空值的前后两个数值来拟合曲线，从而预测空值
    y = s[list(range(n-k,n)) + list(range(n+1,n+1-k))] #取值，range函数返回一个左闭右开（[left,right)）的序列数
    y = y[y.notnull()]#取上一行中取出数值列表中的非空值，保证y的每行都有数值，便于拟合函数
    return lagrange(y.index,list(y))(n) #调用拉格朗日函数，并添加索引

stdScale = 0
if missdealtype == 1:
	ro5mode = df.RO5_violations.dropna().mode().values
	df.RO5_violations[df.RO5_violations.isnull()] = ro5mode
	alogmode = df.AlogP.dropna().mode().values
	df.AlogP[df.AlogP.isnull()] = alogmode
	testdf.RO5_violations[testdf.RO5_violations.isnull()] = ro5mode
	testdf.AlogP[testdf.AlogP.isnull()] = alogmode
elif missdealtype == 2: #mean
	ro5mode = df['RO5_violations'].mean()
	alogmode = df['AlogP'].mean()
	df.RO5_violations[df.RO5_violations.isnull()] = ro5mode
	df.AlogP[df.AlogP.isnull()] = alogmode
	testdf.RO5_violations[testdf.RO5_violations.isnull()] = ro5mode
	testdf.AlogP[testdf.AlogP.isnull()] = alogmode
elif missdealtype == 3: #lagelang
	for i in df.columns: #如果i在data的列名中，data.columns生成的是data的全部列名
		for j in range(len(df)): #len(data)返回了data的长度，若此长度为5，则range(5)会产生从0开始计数的整数列表
			if (df[i].isnull())[j]:#如果data[i][j]为空，则调用函数ployinterp_column为其插值
				df[i][j] = ployinterp_column(df[i],j)
	for i in testdf.columns: #如果i在data的列名中，data.columns生成的是data的全部列名
		for j in range(len(testdf)): #len(data)返回了data的长度，若此长度为5，则range(5)会产生从0开始计数的整数列表
			if (testdf[i].isnull())[j]:#如果data[i][j]为空，则调用函数ployinterp_column为其插值
				testdf[i][j] = ployinterp_column(testdf[i],j)

	
#将后面的特征提取出来，与前面的特征组合
dffeature = df.Features.values
dffeature = np.array(dffeature)
newfeature = []
for line in dffeature:
    line = line[1:-1]
    arra = line.split(',')
    arra = np.array(arra).astype(np.float64)
    newfeature.append(arra)
newfeature = np.array(newfeature)

newfeatureall = np.concatenate((np.array(df.Molecule_max_phase).reshape(df.Molecule_max_phase.shape[0],1),np.array(df['Molecular weight']).reshape(df.Molecule_max_phase.shape[0],1)),axis=1)
newfeatureall = np.concatenate((newfeatureall,np.array(df['RO5_violations']).reshape(df.Molecule_max_phase.shape[0],1)),axis=1)
newfeatureall = np.concatenate((newfeatureall,np.array(df['AlogP']).reshape(df.Molecule_max_phase.shape[0],1)),axis=1)
newfeatureall = np.concatenate((newfeatureall,newfeature),axis=1)
newfeatureall = np.delete(newfeatureall, 5093, 0) # 去掉nan值的记录
newlabel = np.delete(np.array(newlabel), 5093, 0)


#test
testfeature = np.array(testdf.Features.values)
testfeaturearr= []
for line in testfeature:
    line = line[1:-1]
    arra = line.split(',')
    arra = np.array(arra).astype(np.float64)
    testfeaturearr.append(arra)
testfeaturearr = np.array(testfeaturearr)

testfeatureall = np.concatenate((np.array(testdf.Molecule_max_phase).reshape(testdf.Molecule_max_phase.shape[0],1),np.array(testdf['Molecular weight']).reshape(testdf.Molecule_max_phase.shape[0],1)),axis=1)
testfeatureall = np.concatenate((testfeatureall,np.array(testdf['RO5_violations']).reshape(testdf.Molecule_max_phase.shape[0],1)),axis=1)
testfeatureall = np.concatenate((testfeatureall,np.array(testdf['AlogP']).reshape(testdf.Molecule_max_phase.shape[0],1)),axis=1)
testfeatureall = np.concatenate((testfeatureall,testfeaturearr),axis=1)

#过滤没有变化的特征
selectcolumn = (newfeatureall.max(axis=0) != newfeatureall.min(axis=0))
selectcolumn = np.array(selectcolumn)
newfeatureselect = newfeatureall[:,selectcolumn]
testfeatureselect = testfeatureall[:,selectcolumn]

newfeaturenorm = 0
testfeaturenorm = 0
if normaltype == 1:
	stdScale = StandardScaler().fit(newfeatureselect)
	newfeaturenorm = stdScale.transform(newfeatureselect)
	testfeaturenorm = stdScale.transform(testfeatureselect)
elif normaltype == 2:
	newfeature_max = newfeatureselect.max(axis=0)
	newfeaturenorm = newfeatureselect / newfeature_max
	testfeaturenorm = testfeatureselect / newfeature_max
	


from sklearn.linear_model import Ridge
clf = Ridge(alpha=alphacoef)
clf.fit(newfeaturenorm, newlabel)

result = clf.predict(testfeaturenorm)

result

resultfile = open('result.csv','w')
resultfile.write("ID,Label\n")
for index in range(len(testid)):
    result[index] = result[index] if result[index] >= 0 else 0
    resultfile.write(str(testid[index])+","+str(result[index])+"\n")
resultfile.close()