Complementary feature learning across multiple heterogeneous networks and multimodal attribute learning for...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying microRNAs (miRNAs) linked to diseases is crucial for understanding disease pathogenesis.
• Existing methods often struggle to integrate diverse data types and network contexts effectively.

Purpose of the Study:

• To propose CMMDA, a computational method for inferring latent disease-related miRNAs.
• To effectively encode and integrate context relationships, complementary information, and multimodal attributes from heterogeneous networks.

Main Methods:

• Constructed multiple heterogeneous networks based on disease similarities.
• Utilized transformer-based feature representation for miRNA and disease nodes.
• Employed a co-attention fusion mechanism for inter-network information integration.
• Developed a depthwise separable convolution module for multimodal attribute encoding.

Main Results:

• CMMDA demonstrated superior performance in inferring disease-related miRNAs compared to existing methods.
• Ablation studies confirmed the effectiveness of the core innovations within the CMMDA framework.

Conclusions:

• CMMDA offers a powerful approach for miRNA-disease association inference.
• The method's ability to integrate multi-network context and multimodal attributes advances the field of bioinformatics.