School of Information, Yunan Normal University, Kunming, 650092, China.
Key Laboratory of Educational Informatization for Nationalities Ministry of Education, Yunnan Normal University, Kunming, 650092, China.
Sci Rep. 2022 Oct 20;12(1):17519. doi: 10.1038/s41598-022-22447-y.
Identifying lncRNA-disease associations is conducive to the diagnosis, treatment and prevention of diseases. Due to the expensive and time-consuming methods verified by biological experiments, prediction methods based on computational models have gradually become an important means of lncRNA-disease associations discovery. However, existing methods still have challenges to make full use of network topology information to identify potential associations between lncRNA and disease in multi-source data. In this study, we propose a novel method called HGNNLDA for lncRNA-disease association prediction. First, HGNNLDA constructs a heterogeneous network composed of lncRNA similarity network, lncRNA-disease association network and lncRNA-miRNA association network; Then, on this heterogeneous network, various types of strong correlation neighbors with fixed size are sampled for each node by restart random walk; Next, the embedding information of lncRNA and disease in each lncRNA-disease association pair is obtained by the method of type-based neighbor aggregation and all types combination though heterogeneous graph neural network, in which attention mechanism is introduced considering that different types of neighbors will make different contributions to the prediction of lncRNA-disease association. As a result, the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPR) under fivefold cross-validation (5FCV) are 0.9786 and 0.8891, respectively. Compared with five state-of-art prediction models, HGNNLDA has better prediction performance. In addition, in two types of case studies, it is further verified that our method can effectively predict the potential lncRNA-disease associations, and have ability to predict new diseases without any known lncRNAs.
鉴定 lncRNA-疾病关联有助于疾病的诊断、治疗和预防。由于生物实验验证的方法昂贵且耗时，基于计算模型的预测方法逐渐成为发现 lncRNA-疾病关联的重要手段。然而，现有的方法仍然难以充分利用网络拓扑信息，从多源数据中识别 lncRNA 与疾病之间的潜在关联。在这项研究中，我们提出了一种名为 HGNNLDA 的新方法，用于 lncRNA-疾病关联预测。首先，HGNNLDA 构建了一个由 lncRNA 相似性网络、lncRNA-疾病关联网络和 lncRNA-miRNA 关联网络组成的异构网络；然后，在这个异构网络上，通过重启随机游走，对每个节点进行固定大小的各种类型的强关联邻居采样；接下来，通过基于类型的邻居聚合和所有类型组合的方法，通过异构图神经网络获得每个 lncRNA-疾病关联对中 lncRNA 和疾病的嵌入信息，其中考虑到不同类型的邻居对 lncRNA-疾病关联预测的贡献不同，引入了注意力机制。结果，在五重交叉验证（5FCV）下，接收器操作特征曲线（AUC）和精度-召回曲线下面积（AUPR）分别为 0.9786 和 0.8891。与五种最先进的预测模型相比，HGNNLDA 具有更好的预测性能。此外，在两种案例研究中，进一步验证了我们的方法可以有效地预测潜在的 lncRNA-疾病关联，并且具有预测没有任何已知 lncRNAs 的新疾病的能力。
