AttMVGraph: Attention-Based Multimodal Fusion and Variational Graph Learning for SM-miRNA Association Prediction

MiRNA serves as a key noncoding RNA regulating gene expression and is frequently targeted as a therapeutic small molecule (SM). However, relying solely on the experimental identification of SM-miRNA interactions proves costly and inefficient. To address this, this paper proposes an SM-miRNA association prediction method according to attention-based multimodal fusion and variational graph learning (AttMVGraph). The method utilizes Random Walk with Restarts (RWR) topological features and SM/miRNA similarity as multimodal inputs. Adaptive weighted fusion is achieved through feature-enhanced channel attention (FECA), yielding discriminative graph embeddings. Subsequently, a Variational Graph Autoencoder (VGAE) is employed for uncertainty modeling and representation learning. During prediction, a dynamic hard negative mining (DHNM) mechanism is introduced to iteratively select hard negative samples, mitigating extreme positive-negative sample imbalance and strengthening decision boundaries. 5-CV showed that the model produced excellent results, with an AUC = 0.9937 ± 0.0061 (AUC = 0.9727 ± 0.0001) and AUPR = 0.9397 ± 0.0753 (0.8807 ± 0.0589) in Dataset 1 (Dataset 2), verifying the effectiveness and superiority of AttMVGraph. All our data and codes have been uploaded to https://github.com/tr612-maker/AttMVGraph-master.