使用word2vec预训练模型提取DNA/RNA序列特征向量

数据集：https://github.com/alivelxj/m5c-DFRESG/tree/main/m5c/data

代码：

处理数据集

import pandas as pd
import numpy as np
from gensim.models import Word2Vec
from gensim.models.word2vec import LineSentence

def load_data(file):
    seq_sum=[]
    label_sum=[]
    f = open(file).readlines()
    for i in range(0, len(f)-1,2):
        seq = f[i+1].strip('\n')
        seq_sum.append(seq)
        if f[i] == '>+sample\n':
            f[i] = f[i].replace('>+sample\n', '1').strip('\n')
            label_sum.append(int(f[i]))
        if f[i] == '>-sample\n':
            f[i] = f[i].replace('>-sample', '0').strip('\n')
            label_sum.append(int(f[i]))
    return seq_sum, label_sum
X_A_train,y_A_train = load_data(r'm5c-DFRESG-main\m5c\data\Athaliana\Arabidopsis_train.fasta')
X_A_ind, y_A_ind =load_data(r'm5c-DFRESG-main\m5c\data\Athaliana\Arabidopsis_indep.fasta')
y_A_ind=np.array(y_A_ind)
y_A_ind=y_A_ind.reshape(-1,1)

模型部分+训练+提取特征

class word2vec(object):
    def __init__(self, seq, test_seq, k, num_feature, min_count, num_workers, context, seed, epoch, model_name):
        self.seq = seq
        self.test_seq = test_seq
        self.k = k
        self.num_feature = num_feature
        self.min_count = min_count
        self.num_works= num_workers
        self.context = context
        self.seed = seed
        self.epoch = epoch
        self.model_name = model_name

    '''  通过调用gensim.word2vec来提取词向量
    seq:构建语料库所需的数据
    test_seq:
    k:使用k-mer分词，k:int  1,2,3....
    num_feature : 最终词向量的维度
    min_count: 可以对字典做截断. 词频少于min_count次数的单词会被丢弃掉, 默认值为5
    num_workers :参数控制训练的并行数
    context:表示当前词与预测词在一个句子中的最大距离是多少
    seed:用于随机数发生器。与初始化词向量有关
    epoch:迭代次数
    model_name:训练好后的模型文件
    '''
    #分词
    def seperate_word(self,single_seq):
        sentence = ''
        for i in range(len(single_seq)-self.k+1):
            sentence += single_seq[i:i+self.k]+ " "
        sentence = sentence[0:len(sentence)-1]
        return sentence

    def process_all_seq(self):

        seq_list = []
        for i in range(len(self.seq)):
            i = self.seq[i].strip('\n')
            line = self.seperate_word(i)
            # line = line+'\n'
            seq_list.append(line)

        return seq_list

    # 如果需要重新构建词汇表，需要执行下面函数
    # def covert_file(self):
    #     seq = open('train', 'w')
    #     for i in range(len(self.process_all_seq())):
    #         corpus = self.process_all_seq()[i].strip('')
    #         seq.write(corpus+'\n')
    #     seq.close()
    #构建词汇表之后，利用词汇表训练word2vec模型并保存
    def word_model(self):
        sentence = LineSentence('train', max_sentence_length=15000)
        model = Word2Vec(sentence, workers=self.num_works, vector_size=self.num_feature, min_count=self.min_count,
                         window=self.context, seed=self.seed, epochs=self.epoch)
        model.init_sims(replace=False)
        model.save("train_model")

        return model
# 需要训练的数据进行分词
    def getseq_split(self):
        seq_list = []
        for i in self.test_seq:
            seq_list.append(self.seperate_word(i).split(" "))
        return seq_list

    # def process_test_seq(self):
    #
    #     seq_list = []
    #     for i in range(len(self.test_seq)):
    #         i = self.seq[i].strip('\n')
    #         line = self.seperate_word(i).strip('')
    #         # line = line+'\n'
    #         seq_list.append(line)
    #
    #     return seq_list

    # def covert_test(self):
    #     seq = open('test', 'w')
    #     for i in range(len(self.process_test_seq())):
    #         corpus = self.process_test_seq()[i].strip('')
    #         seq.write(corpus+'\n')
    #     seq.close()

    # def concatenate(self,pos_seq,neg_seq,total_seq):
    #     pos_file = open(pos_seq)
    #     neg_file = open(neg_seq)
    #     total_file = open(total_seq,'w')
    #     for i in pos_file:
    #         if '>' not in i:
    #             total_file.write('1\t'+i)
    #     for i in neg_file:
    #         if '>' not in i:
    #             total_file.write('0\t'+i)
    #     pos_file.close()
    #     neg_file.close()
    #     total_file.close()
    # 返回特征矩阵
    def get_feature_matrix(self):
        test_data = self.getseq_split()
        print(len(test_data))
        feature_matrix = []
        for i in test_data:

            feature_matrix.append(np.mean(Word2Vec.load(self.model_name).wv[i],axis=0))

        return feature_matrix

    def concat_feature_label(self, label_file):
        feature = self.get_feature_matrix()
        label = label_file
        feature_file = np.concatenate([feature, label], axis=1)
        return feature_file

调用

if __name__=='__main__':
     test =word2vec(X_A_train,X_A_ind, 2, 100, 1, 10, 20, 42, 10, 'train_model')
     a=test.concat_feature_label(y_A_ind)
     pd.DataFrame(a).to_csv("feature.csv")

总结：最终会生成一个shape为（seq_length, num_feature) 的数组

不足：语料库比较小，可能效果较差