CAPTLDA：基于胶囊网络和Transformer预测LncRNA-疾病关联

CAPTLDA: Predicting LncRNA-Disease Associations Based on Capsule Networks and Transformers

摘要: 长链非编码RNA (lncRNA)是一类长度超过200个核苷酸的转录物，在多种疾病的发病机制中发挥关键作用。因此，阐明lncRNA与疾病之间的关联对于理解潜在的发病机制和开发新的疾病预防、诊断和治疗策略至关重要。虽然传统的生物学实验对于预测长链非编码RNA-疾病关联(LDA)是有价值的，但往往费用高昂且耗时。开发有效的LDA预测计算模型是有必要的。当前的计算方法在有效整合多源数据和捕获异质生物网络中的复杂高阶关系模式方面经常遇到限制。这项研究提出了一种新的计算框架命名为CAPTLDA，将lncRNA、疾病和miRNA的相似性和关联整合到一个加权的异构网络邻接矩阵中，引入了胶囊网络，以增强特征学习。此外，还采用Transformer编码器，它结合了全局多头代理注意力机制和并行的多头局部注意力机制，以全面捕获全局依赖关系和局部上下文信息，最终实现准确的LDA预测。在两个基准数据集上进行的综合计算实验表明，模型在性能上优于先进的现有方法。案例研究进一步验证了它在识别潜在疾病相关lncRNA方面的有效性。

Abstract: Long non-coding RNAs (lncRNAs) are transcripts exceeding 200 nucleotides in length and play pivotal roles in the pathogenesis of various diseases. Therefore, elucidating the associations between lncRNAs and diseases is crucial for understanding underlying mechanisms and developing novel strategies for disease prevention, diagnosis, and treatment. While traditional biological experiments are valuable for predicting lncRNA-disease associations (LDA), they are often costly and time-consuming. Developing effective computational models for LDA prediction is therefore necessary. Current computational methods frequently encounter limitations in effectively integrating multi-source data and capturing complex higher-order relationship patterns in heterogeneous biological networks. This study proposes a novel computational framework named CAPTLDA, which integrates the similarities and associations of lncRNAs, diseases, and miRNAs into a weighted heterogeneous network adjacency matrix. A capsule network is introduced to enhance feature learning. Additionally, a Transformer encoder is employed, combining a global multi-head agent attention mechanism with parallel multi-head local attention mechanisms to comprehensively capture global dependencies and local contextual information, ultimately achieving accurate LDA prediction. Comprehensive computational experiments on two benchmark datasets demonstrate that our model outperforms existing advanced methods in performance. Case studies further validate its effectiveness in identifying potential disease-related lncRNAs.