Existing circRNA-miRNA interaction prediction methods have not fully exploited the spatial folding information in circRNAs and pre-miRNAs. Furthermore, current methods inadequately address intramolecular modal consistency and intermolecular discrimination across distinct modalities, leading to suboptimal discriminative performance. Traditional attention mechanisms are computationally intensive and lack fine-grained, balanced modal weight allocation, impeding efficient feature fusion. To address these limitations, we propose GSSCMI, a novel method comprising three key components. First, an information integration module incorporates similarity information, sequence features, and secondary structure features. Second, a multimodal contrastive learning module processes features across three modalities, enhancing both intra-modal consistency for individual circRNA/miRNA and inter-modal discrimination between different circRNAs/miRNAs. Third, we innovatively design an efficient co-attention mechanism that simultaneously modulates fused modalities through unified and fine-grained attention scores, achieving balanced fusion while significantly reducing computational overhead. Experimental results demonstrate GSSCMI outperforms existing methods, with improvements of 9.53% in MCC and 6.40% in F1. Ablation studies further show the efficient co-attention reduces convergence iterations by approximately 58%. Compared to the initial co-attention, it reduces computational cost by 42.94% while improving MCC by 13.32% and ACC by 6.69%. Additionally, we identified regulatory sites through structure-based A-to-I editing and elucidated sequence-level inter-token dependencies via attention visualization.
Lihao Sun, Xin Wang, Fang Wang et al.· IEEE journal of biomedical a...· 0 citations
Abstract Motivation Circular RNAs (circRNAs) interact with microRNAs (miRNAs) to regulate gene expression and influence disease progression. However, traditional models tend to overlook the significant contributions of certain features when dealing with diverse sequence information, resulting in the inability to capture some deep topological structures and thus leaving room for improvement in prediction performance. Results We propose HMA-GCA, a novel framework that integrates hybrid manifold augmentation and gated cross-attention for CMI prediction. The model first constructs multi-scale descriptors by combining sequence-derived features (K-mer, CTD, Doc2Vec) and topological features (Role2Vec, node degree, neighborhood proximity). It then applies PCA for global linear projection and UMAP for local nonlinear manifold learning, enhancing feature representations while preserving intrinsic data geometry. A channel-wise gated cross-attention mechanism dynamically controls the injection of miRNA information into circRNA representations. Extensive experiments on three benchmark datasets show that HMA-GCA consistently outperforms state-of-the-art methods across multiple metrics. To ensure interpretability, we conducted SHAP analysis to quantify the contribution of each feature type, revealing that sequence-derived features and topological similarities are the most influential. Ablation studies confirm the necessity of each module, while case studies demonstrate that top-ranked predictions are supported by literature evidence. Overall, HMA-GCA not only achieves state-of-the-art predictive performance but also provides interpretable insights into the molecular features. Availability and implementation The source code and data are freely available at https://github.com/Lixunwind/Prediction-circ-mi-by-Gate.git. The implementation is based on Python and the required dependencies are listed in the repository.
Yun-peng Hu, Yansu Wang, Yifeng Bai et al.· Bioinform.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
