CCLDA: prediction of lncRNA-disease associations based on Convolutional Block Attention Module and Capsule Network

Several studies have shown that long non-coding RNAs (lncRNAs) influence the biological processes of many diseases, including disease onset, progression, and recovery. Therefore, predicting potential lncRNA-disease associations (LDAs) is crucial for enhancing disease diagnosis and therapy. Compared with biological experimental methods for identifying potential LDAs, computational approaches offer advantages in terms of efficiency and cost-effectiveness. In this study, we introduce a novel deep learning approach, CCLDA, for predicting LDAs. First, we constructed the functional similarity matrix, Gaussian similarity matrix, and sequence similarity matrix for lncRNAs, and the semantic similarity matrix and Gaussian similarity matrix for diseases, applying a matrix fusion process. Then, lncRNA-disease pairs were constructed, and feature extraction was conducted using a multilayer autoencoder (AE). The extracted features were then fed into a capsule network to train the model and generate prediction scores. A Convolutional Attention Module (CBAM) was integrated into the capsule network to assign weights to convolved features in both channel and spatial dimensions, enhancing overall model prediction performance. We compared CCLDA with other models on two datasets, and the results demonstrated that CCLDA outperformed existing LDA prediction methods. Ablation experiments further confirmed the necessity of CCLDA's components, and case studies on both datasets indicated that CCLDA holds significant potential for predicting novel LDAs. Based on these results, CCLDA will be important for research in lncRNA-disease prediction related fields.