HNMDA：异构网络驱动的miRNA-疾病关联预测

需积分: 50 36 浏览量更新于2024-08-12 3 收藏 1.64MB PDF 举报

"这篇科研文章探讨了一种名为HNMDA（Heterogeneous Network-based MiRNA-Disease Association prediction）的计算模型，该模型用于预测microRNA（miRNA）与疾病之间的关联。miRNA在人类多种疾病的发生和发展中起着关键作用，但实验验证这些关联既昂贵又耗时。HNMDA模型结合了已知的miRNA-疾病关联、miRNA功能相似性、疾病语义相似性和高斯交互轮廓内核相似度，通过异构网络的方法进行预测。在交叉验证中，HNMDA表现出优于其他方法的预测性能，AUC达到0.8394。在针对乳腺肿瘤、食道肿瘤和肾肿瘤的案例研究中，预测的miRNA中有82%、76%和84%分别被证实与这些疾病相关。此外，HNMDA在新疾病及使用不同已知关联数据库的案例中，也显示了高比例的预测准确性。" HNMDA模型是建立在异构网络的基础上，异构网络允许不同类型的节点（如miRNA、疾病和它们的属性）和边（如相似性关系）的整合。在这个模型中，首先利用网络扩散和重启的随机游走算法，模拟信息在网络中的传播和收敛过程。接着，通过找到最佳投影，将miRNA空间映射到疾病空间，以预测尚未被实验验证的关联。高斯交互轮廓内核相似度是一种强大的工具，能弥补传统相似度计算方法的局限，更好地捕捉节点之间的复杂关系。在评估中，HNMDA不仅在交叉验证中表现优秀，而且在实际的案例研究中也证明了其预测能力。对于HMDD V2.0数据库中记录的乳腺肿瘤、食道肿瘤和肾肿瘤，前50个预测的miRNA关联中有相当大的比例被证实。即使在没有已知关联的新疾病中，以及使用HMDD V1.0数据库时，HNMDA的预测结果也有很高的验证比例，这体现了模型的泛化能力和实用性。 HNMDA模型提供了一种有效且准确的方法，能够辅助研究人员预测潜在的miRNA-疾病关联，从而加速疾病机制的理解和新疗法的开发。这种方法有望在未来的生物信息学研究和临床实践中发挥重要作用。

985Molecular Genetics and Genomics (2018) 293:983–995

1 3

model which is Restricted Boltzmann Machine for Multiple

types of MiRNA–Disease Association prediction (RBM-

MMDA) based on known miRNA–disease associations.

Although RBMMMDA performed well on both predicting

miRNA–disease associations and miRNA–disease asso-

ciation types, the choice of parameters in the model is still

unsolved. Li etal. (2017) developed a model called Matrix

Completion for MiRNA–Disease Association prediction

(MCMDA) using singular value thresholding (SVT) algo-

rithm to complete the miRNA–disease association matrix.

However, this model cannot work for diseases with no

known related miRNAs. Chen etal. further developed a

model, which is Ranking-based KNN for MiRNA–Disease

Association prediction (RKNNMDA). In this model, an ini-

tial KNN-based ranking method was ﬁrst applied. Due to

biases caused by the drawback of KNN, SVM is introduced

to re-rank the previous ranked neighbors. Although SVM

is introduced to the model, bias might still exist in the ﬁnal

scores. Besides, an ideal method to combine KNN, SVM and

weighted voting is still needed (Chen etal. 2017).

To further exploit the potential associations between miR-

NAs and diseases, researchers have proposed deep learn-

ing methods (Chen etal. 2017; Fu and Peng 2017). Chen

etal. has introduced DRMDA (Deep Representations-based

MiRNA–Disease Association Prediction), using stacked

auto-encoder to obtain the abstract representations of the

raw data. A SVM classiﬁer is stacked on the top of the auto-

encoder. However, since the need of negative data, SVM

classiﬁer can not perform as good as expected. The param-

eters in DRMDA are also not easy to optimize. Fu etal.

also proposed an auto-encoder-based method to predict

miRNA–disease associations. They fed miRNA–miRNA

similarity network and disease–disease similarity network

into stacked auto-encoders, respectively, to extract features

from both similarity networks. The extracted features were

then concatenated as combined features and fed into a three-

layer fully connected network to calculate the probability of

the miRNA and disease being associated.

In this study, we proposed a network integration approach

called Heterogeneous Network-based MiRNA–Disease

Association prediction (HNMDA) to predict potential

miRNA–disease associations. To obtain a high accuracy,

we combined the miRNA similarity and disease semantic

similarity with Gaussian interaction proﬁle kernel simi-

larities. We ﬁrst built up miRNA and disease similarity

networks, based on the similarity data, respectively. Then

we introduced a network diﬀusion algorithm called Ran-

dom Walk with Restart (RWR), with which we can take the

global structure of the networks into consideration. Finally,

we managed to ﬁnd an optimal projection from the miRNA

space onto the disease space, which enabled the predic-

tion of potential miRNA–disease associations according to

the geometric proximity of the mapped vectors. To get the

optimal projection, we turned it into an alternating minimi-

zation problem, and applied an inductive matrix comple-

tion method to solve it (Natarajan and Dhillon 2014). This

method worked well for diseases without any known related

miRNAs. Furthermore, leave-one-out cross-validation

(LOOCV) was introduced to evaluate our model. The AUC

of LOOCV is 0.8394. Moreover, we evaluated HNMDA

with three kinds of case studies. In the ﬁrst case, we tested

our model on breast neoplasms, esophageal neoplasms and

kidney neoplasms, and there were 41, 38 and 42 out of top

50 miRNAs conﬁrmed by experiments, respectively. In the

second case, we applied HNMDA on the test diseases whose

known associations with miRNAs were set to be unknown

ones. As a result, 49 out of top 50 miRNAs predicted to be

associated with hepatocellular carcinoma were experimen-

tally veriﬁed. And to test the robustness of HNMDA, we

tested our model using HMDD V1.0 database, and 40 out

of top 50 potential lymphoma-related miRNAs were experi-

mentally conﬁrmed. Therefore, HNMDA is proved to be an

accurate and eﬀective method in predicting miRNA–disease

associations.

Results

Performance evaluation

LOOCV was introduced to evaluate the accuracy of

HNMDA. Through the process of LOOCV, we left out each

known miRNA–disease association in turn as test sample,

and other known miRNA–disease associations were used for

training. Those miRNA–disease pairs which have no con-

ﬁrmed associations were taken to be candidate pairs. We

compared the score of each test sample with scores of all the

candidate pairs, and if its rank was above the threshold given

in advance, it will be considered as a successful prediction.

To further evaluate HNMDA, we drew Receiver operating

characteristics (ROC) curve by plotting true-positive rate

(TRP, sensitivity) versus the false-positive rate (FPR, 1−sen-

sitivity) at diﬀerent thresholds. Sensitivity means the ratio of

the positive samples correctly predicted among all positives

ones. And speciﬁcity refers to the ratio of negative samples

correctly predicted among all negative ones. Area under the

ROC curve (AUC) is calculated to indicate the prediction

ability of HNMDA. AUC = 0.5 indicates the model only has

a random performance, while AUC = 1 indicates the model

performs perfectly in the prediction of miRNA–disease

associations. Compared with RKNNMDA and WBSMDA

of which the AUCs are 0.7159 and 0.8030, respectively,

HNMDA got the AUC of 0.8394, which has improved the

accuracy of predicting potential miRNA–disease association

(see Fig.1).

剩余12页未读，继续阅读

weixin_38713167

粉丝: 6
资源: 938

HNMDA：异构网络驱动的miRNA-疾病关联预测

数据融合matlab代码-SNFHGILMI-master:SNFHGILMI大师

基于元路径的MiRNA-疾病关联预测

L1-norm-Graph:miRNA-疾病关联预测的新型半监督方法

数据融合matlab代码-BIRWMDA:融合多个生物网络以有效预测miRNA-疾病关联

神经诱导图卷积网络：miRNA-疾病关联预测的新方法

PBMDA：新型miRNA-疾病关联预测计算模型

NDAMDA：利用网络距离预测miRNA-疾病关联的高效方法

NARRMDA：一种新的预测miRNA-疾病关联的算法

详细讲讲神经网络在miRNA-药物关联预测的国内外研究现状

详细讲讲miRNA-药物关联预测的国内外研究现状

最新资源