MOMHCA-SG: a multi-head cross-attention and similar graph convolutional network framework for Alzheimer's disease...

Area of Science:

• Computational biology
• Neuroscience
• Genomics

Background:

• Alzheimer's disease (AD) presents complex, heterogeneous characteristics, challenging early detection and understanding.
• Conventional unimodal approaches struggle to capture intricate interactions across biological layers in AD.

Purpose of the Study:

• To develop a novel multi-omics integration framework, MOMHCA-SG, for enhanced Alzheimer's disease research.
• To improve cell-type classification and identify key molecular players in AD pathogenesis.

Main Methods:

• The MOMHCA-SG framework integrates multi-head cross-attention (MHCA) and graph convolutional networks (GCN) with similarity network fusion (SNF).
• It employs autoencoders for dimensionality reduction, MHCA for inter-omic dependencies, contrastive learning for representation refinement, and GCN for classification using scRNA-seq and scATAC-seq data.
• The framework processes high-dimensional omics data while preserving feature integrity and dynamically integrating heterogeneous information.

Main Results:

• MOMHCA-SG achieved superior performance in data integration and cell-type classification tasks on AD datasets, with high scores for Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and Adjusted Mutual Information (AMI).
• Classification accuracy exceeded 0.98, demonstrating the framework's effectiveness.
• Downstream bioinformatics analyses identified significant genes and signaling pathways implicated in AD pathogenesis, confirming biological interpretability.

Conclusions:

• MOMHCA-SG effectively integrates heterogeneous omics data and captures cross-modality relationships, offering a powerful tool for Alzheimer's disease research.
• The framework advances precision medicine by enabling a deeper understanding of neurodegenerative disease mechanisms.