预测与疾病相关的候选长非编码RNA（lncRNA）对于探索疾病发病机制具有益处，因为lncRNA与人类疾病的发生和发展息息相关。充分提取每个个体lncRNA网络和个体疾病网络中的特定和局部拓扑结构，并整合连接关系的信息是一个长期而有挑战性的任务。我们提出了一种新的基于图学习的预测方法，来编码每个网络的特定和局部拓扑结构，具有不同连接关系的邻居拓扑结构和成对属性。首先，我们构建了一个包含所有lncRNA节点及其相似性的lncRNA网络，以及一个包含所有疾病节点和疾病相似性的单一疾病网络。然后，构建了一个异构网络，用于嵌入所有lncRNA、疾病和miRNA节点及其各种连接。随后，设计了一种连接敏感的图神经网络，深度整合异构网络中的邻居节点属性和连接特征，并学习邻居拓扑表示。我们还构建了连接级和拓扑表示级的注意机制，以提取有信息量的连接和拓扑表示。最后，我们构建了一个带有加权残差的多层卷积神经网络，自适应地补充详细特征到成对属性编码中。综合实验和比较结果表明，NCPred优于七种先进的预测方法。消融研究证明了局部拓扑学习、邻居拓扑学习和成对属性编码的重要性。前列腺癌、肺癌和乳腺癌的案例研究进一步揭示了NCPred在筛选潜在候选与疾病相关的lncRNA方面的能力。版权所有©2023 Elsevier Ltd. 保留所有权利。

Predicting disease-related candidate long noncoding RNAs (lncRNAs) is beneficial for exploring disease pathogenesis due to the close relations between lncRNAs and the occurrence and development of human diseases. It is a long-term and challenging task to adequately extract specific and local topologies in individual lncRNA network and individual disease network, and integrate the information of the connection relationships. We propose a new graph learning-based prediction method to encode specific and local topologies from each individual network, neighbor topologies with different connection relationships, and pairwise attributes. We first construct a lncRNA network composed of all the lncRNA nodes and their similarities, and a single disease network that contains all the disease nodes and disease similarities. Then, a network-aware graph convolutional autoencoder is constructed to encode the specific and local topologies of each network. Secondly, a heterogeneous network is established to embed all lncRNA, disease, and miRNA nodes and their various connections. Afterwards, a connection-sensitive graph neural network is designed to deeply integrate the neighbor node attributes and connection characteristics in the heterogeneous network and learn neighbor topological representations. We also construct both connection-level and topology representation-level attention mechanisms to extract informative connections and topological representations. Finally, we build a multi-layer convolutional neural networks with weighted residuals to adaptively complement the detailed features to pairwise attribute encoding. Comprehensive experiments and comparison results demonstrated that NCPred outperforms seven state-of-the-art prediction methods. The ablation studies demonstrated the importance of local topology learning, neighbor topology learning, and pairwise attribute encoding. Case studies on prostate, lung, and breast cancers further revealed NCPred's capacity to screen potential candidate disease-related lncRNAs.Copyright © 2023 Elsevier Ltd. All rights reserved.