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

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying microRNA-disease relationships is crucial for understanding disease mechanisms and developing therapies.
• Current computational models often fail to leverage heterogeneous biological data and capture complex non-linear associations between microRNAs and diseases.

Purpose of the Study:

• To develop an innovative computational framework, BADMDA, for accurately identifying potential microRNA-disease associations.
• To overcome limitations of existing models by effectively integrating diverse biological resources and capturing intricate non-linear dependencies.

Main Methods:

• BADMDA combines centered kernel alignment-based multiple kernel learning (CKA-MKL) for adaptive similarity kernel weighting.
• A bipartite attributed graph with drop aggregation (DropAGG) is used for expressive multi-level feature learning and mitigating over-smoothing.
• A bi-directional attention mechanism captures reciprocal dependencies, enhancing feature quality for association inference.

Main Results:

• BADMDA achieved superior accuracy in predicting microRNA-disease associations on HMDD v2.0 and HMDD v3.2 datasets.
• The framework demonstrated high performance with AUC values of 0.9372 and 0.9533, and AUPR values of 0.9348 and 0.9525, respectively.
• Case studies on lung neoplasms and hepatocellular carcinoma validated the robustness and biological relevance of the predicted associations.

Conclusions:

• BADMDA effectively integrates heterogeneous biological similarities and retains complex non-linear dependencies through attention-driven feature learning.
• The framework represents a significant advancement in computational approaches for identifying microRNA-disease associations.
• BADMDA's performance surpasses existing state-of-the-art methods, offering a promising tool for biomedical research and therapeutic development.