Bi-directional attention network with drop aggregation for microRNA-disease association prediction

Identifying the potential relationships between microRNAs (miRNAs) and human diseases is vital for advancing our knowledge of pathological mechanisms and promoting therapeutic development. Nevertheless, many current computational models make insufficient use of heterogeneous biological resources and depend on simplistic feature extraction, which limits their ability to capture the intricate non-linear associations between miRNAs and diseases. To overcome these challenges, we present BADMDA, an innovative framework that combines centered kernel alignment-based multiple kernel learning (CKA-MKL), drop aggregation (DropAGG), and a bi-directional attention mechanism. In this design, CKA-MKL adaptively assigns weights to different similarity kernels, thereby achieving an effective integration of diverse information sources. Based on established miRNA-disease associations (MDAs), we then build a bipartite attributed graph, where DropAGG learns expressive multi-level features and alleviates over-smoothing. The bi-directional attention component further improves feature quality by capturing reciprocal dependencies between diseases and miRNAs. A multi-layer perceptron is finally employed to infer novel associations from the enhanced representations. The superior accuracy of BADMDA is attributed to its ability to adaptively fuse heterogeneous similarities and retain complex non-linear dependencies through attention-driven feature learning. On HMDD v2.0 and HMDD v3.2, it achieves AUC values of 0.9372 and 0.9533 and AUPR values of 0.9348 and 0.9525, respectively, surpassing ten advanced methods. Case studies on lung neoplasms and hepatocellular carcinoma further confirm the robustness and biological relevance of BADMDA.