AMFCL: Predicting miRNA-Disease Associations Through Adaptive Multi-source Modality Fusion and Contrastive Learning

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNA (miRNA) dysregulation is implicated in disease progression.
• Identifying miRNA-disease associations (MDAs) is crucial for biomarker discovery.
• Computational methods offer rapid and cost-effective alternatives to traditional biological approaches for MDA identification.

Purpose of the Study:

• To develop a novel computational model, AMFCL, for accurate and efficient miRNA-disease association prediction.
• To address key challenges in existing computational methods, including multi-source information integration, feature fusion optimization, and graph-based model over-smoothing.

Main Methods:

• Construction of three network types to represent miRNA-disease relationships.
• Utilizing multi-layer graph sample and aggregate (GraphSAGE) for learning node representations.
• Employing an adaptive fusion mechanism (AFM) for dynamic feature representation weighting.
• Incorporating residual connections to mitigate over-smoothing in graph-based models.
• Applying contrastive learning (CL) to enhance the robustness of miRNA and disease embeddings.
• Using a multi-layer perceptron (MLP) for computing MDA scores.

Main Results:

• AMFCL demonstrated significant improvements in MDA prediction accuracy compared to existing advanced models.
• Experimental results validated the effectiveness of AMFCL in identifying novel miRNA-disease associations.
• Case studies further confirmed the practical utility and robustness of the proposed approach.

Conclusions:

• AMFCL offers a powerful and effective computational framework for uncovering miRNA-disease associations.
• The model's ability to integrate multi-source information and optimize feature fusion contributes to enhanced biomarker discovery.
• AMFCL shows promise for advancing precision medicine through improved understanding of disease mechanisms.