大规模miRNA-lncRNA交互网络的新型链接预测方法 - CSDN文库

90 浏览量更新于2024-08-27 收藏 1.92MB PDF 举报

"这篇研究论文探讨了在大规模的miRNA-lncRNA交互网络中的新颖链接预测方法，基于二分图的框架。当前对于miRNA-lncRNA相互作用的认知仍然有限，预测miRNA的目标lncRNA的努力相对较少。研究发现lncRNA和miRNA之间的交互模式与相对表达水平密切相关，形成了一种调控机制。通过分析编码非编码共表达网络和序列数据，可以度量miRNA和lncRNA之间的相似性。研究中提出了两个关键发现：(i) lncRNA/miRNA倾向于与具有相似表达谱的miRNA/lncRNA协同交互，反之亦然；(ii)那些与一组共同目标基因相互作用的miRNA往往共同靶向这些基因。" 这篇研究论文属于生物信息学领域，具体关注的是miRNA（微小核糖核酸）与lncRNA（长链非编码RNA）之间的相互作用。miRNA是一类小分子RNA，广泛参与基因表达的调控，而lncRNA虽然不编码蛋白质，但对基因表达和细胞功能有重要影响。在当前的研究背景下，miRNA-lncRNA交互网络的全面理解是至关重要的，因为它们在多种生物学过程中发挥着关键作用，包括疾病的发生和发展。研究中提到的“链接预测”是指在已知的交互网络基础上，利用各种算法和统计模型预测未被发现的miRNA与lncRNA之间的相互作用。这通常涉及到特征提取，例如miRNA和lncRNA的表达水平、序列相似性和共表达模式。通过对这些特征进行数学分析，研究者发现了两个关键现象： 1. 相似表达模式的lncRNA和miRNA倾向于相互作用：这意味着在细胞或组织中表达水平相近的lncRNA和miRNA更可能形成相互作用对。这一发现为理解miRNA-lncRNA调控网络的动态性质提供了新视角，有助于预测新的交互关系。 2. 共同目标基因的miRNA常常一起发挥作用：这种现象表明，某些miRNA可能共享相似的靶基因集合，这可能是因为它们共同参与了特定的生物学过程或者有类似的调控功能。这一发现对于解析miRNA的调控网络以及在疾病中的潜在作用具有重要意义。通过这些发现，研究者提出了一种新的方法来预测大规模miRNA-lncRNA交互网络中的未知链接，这将有助于扩大我们对非编码RNA调控网络的理解，从而推动相关疾病的早期诊断和治疗策略的发展。此外，这项工作也强调了数据分析和数学建模在揭示复杂生物系统内在规律中的重要作用。

inference of ceRNA interactions mediated via phosphatase

and tensin homolog (PTEN). However, all the factors used

by this model, such as degradation and transcription rates

for association and dissociation of miRNA/ceRNAs com-

plexes [24], are too difficult to be surveyed for most of miR-

NAs and lncRNAs. Therefore, it is not feasible to

extensively use this kinetic model for the inference of

miRNA-lncRNA interactions. Increasing evidences [30, 31]

demonstrated that lncRNAs are also presumably

co-regulated in expression networks, and multiple lncRNAs

could involve in the biological regulation processes by syn-

ergistically interacting particular miRNA clusters. Accord-

ingly, the expression pattern of lncRNA-lncRNA synergistic

network has recently attracted increasing attention.

In this work, we develop a group-preference Bayesian

collaborative filtering model called GBCF to pick up a

top-k probability ranking list for an individual miRNA or

lncRNA based on the known miRNA-lncRNA interaction

network derived from lncRNASNP database. Since the

known miRNA-lncRNA interactions in the lncRNASNP

database are all positive, the negative samples are relatively

hard to be collected. This prediction task is actually a

semi-supervised one only treating the known interactions

as positive samples. The semi-supervised prediction task

can properly utilize enough side information beneficial for

the prediction performance. Particularly, we first propose

the local scoring scheme to alleviate the prediction prefer-

ence caused by the disproportion of the known

miRNA-lncRNA interaction network. In this scoring sys-

tem, we implemented both leave-one-out cross validation

(LOOCV) and k-fold cross validation to evaluate the pre-

diction performance of the proposed model. The experi-

mental result demonstrated that GBCF obtain the reliable

prediction performance and achieve the higher AUC (area

under ROC curve) of 0.9193 compared with a few repre-

sentative classical classifiers and the state-of-the-art model

EPLMI [32]. GBCF obtained the average AUCs of 0.8354

+/− 0.0079, 0.8615+/− 0.0078 and 0.8928+/− 0.0082 in the

frameworks of 2-fold, 5-fold and 10-fold cross validations,

respectively. To better describe the similarities among

miRNAs and lncRNAs, we leveraged three diverse types

of biological information, i.e., expression profile,

coding-non-coding co-expression networks and sequence

data. Using a series of 5-fold cross validations and correl-

ation analysis of RNA clusters, the experimental compari-

son demonstrated that the miRNA and lncRNA similarity

should be measured by the biological function-based and

expression profile-based correlations, respectively.

Results

The experiment result in cross validations

Using LOOCV, we compared GBCF with a few classical

classifiers including [33–36] as well as the state-of-the-art

model EPLMI [30] as baseline. Note that, all the compared

models were built on the same information source as

GBCF. EPLMI is a two-way diffusion model first proposed

for the prediction of large-scale miRNA-lncRNA interac-

tions. Unlike GBCF, EPLMI adopts a global scoring

scheme to rank the most potential novel miRNA-lncRNA

interactions among all unobserved samples. We also tried

to explore the potential of these classical classifiers from

different perspectives. For example, Katz can be catego-

rized as the network-based measurement method by cal-

culating the nodes’ similarity in a bipartite graph.

Singular-value decomposition (SVD) is used to decompose

the known interaction network into three relatively smaller

matrices for construction of probability matrix. Latent fac-

tor model (LFM) aims to explain observed associations in

terms of two latent factors (also called hidden variables),

which are iteratively optimized for matrix product as prob-

ability matrix. Since GBCF model adopts a specific

group-preference Bayesian collaborative filtering (CF)

technique, we also compared it with typical lncRNA-based

and miRNA-based CF models, respectively.

The performance comparison via LOOCV is shown in

Fig. 1. Among these models, GBCF achieves the best pre-

diction performance with the highest AUC value of

0.9193. The miRNA-based CF, lncRNA-based CF, EPLMI,

SVD-based model and basic LFM obtain the AUC values

of 0.9089, 0.8880, 0,8847, 0.8402 and 0.8680 respectively.

It is noteworthy that the CF-based models seems to per-

form better than others do. This phenomenon could be at-

tributed to their capability of automatic collecting

extrinsic preferences from other RNAs. Although EPLMI

model still maintain reasonable prediction accuracy, the

local ranking scheme limit its performance to a certain ex-

tent. GBCF is developed from the previous approach of

the recommended system, it is more efficient to deal with

the sparse dataset than EPLMI. In a word, the LOOCV re-

sults demonstrate the reliability of GBCF.

Insufficient training samples would greatly affect the

prediction accuracy (sparsity = 2.49%). To evaluate the

performance of GBCF in terms of diverse sparsity,

2-fold, 5-fold and 10-fold cross validations were con-

ducted, respectively. As shown in Table 1, GBCF model

achieves the average AUCs of 0.8354+/− 0.0079 when

the number of training samples drops to a half. In

addition, the result suggests that GBCF model shows a

strong ro bustness to different level of training data spar-

sities. We also used 5-fold cross validation to assess the

performance of GBCF with lncRNA-based group prefer-

ence instead. The average AUCs of 0.8612+/− 0.0080 ob-

tained suggest that miRNA- and lncRNA- based group

preferences contribute equally to the prediction per-

formance of GBCF. Considering the complex competi-

tion mechanisms in ceRNA network and the lack of

investigation into the competition patterns for sequester-

ing miRNAs, we provided the top-50 ranking lists of

Huang et al. BMC Medical Genomics 2018, 11(Suppl 6):113 Page 19 of 112

剩余10页未读，继续阅读

weixin_38662327

粉丝: 5
资源: 922

最新资源