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

Introduction: Alzheimer's disease (AD) is a complex neurodegenerative disorder with di-verse cellular and molecular characteristics. Due to its heterogeneous nature, early detection and understanding of AD are challenging, and conventional unimodal approaches often fail to capture the intricate interactions across different biological layers. Methods: We propose MOMHCA-SG, a novel multi-omics integration framework that combines a multi-head cross-attention mechanism (MHCA) with a graph convolutional network (GCN) guided by similarity network fusion (SNF). The framework first employs an autoencoder (AE) for dimensionality reduction and feature extraction, followed by MHCA to model inter-omic dependencies. A contrastive learning strategy is then utilized to refine latent representations, while the GCN performs final cell-type classification using fused similarity networks derived from both scRNA-seq and scATAC-seq datasets related to AD. Results: Extensive experiments on publicly available AD datasets demonstrate that MOMHCA-SG attains superior performance in both integration and classification tasks, achieving an ARI of 0.99, NMI of 0.98, and AMI of 0.98, with classification accuracy exceeding 0.98. Downstream bioinformatics analyses further identify key genes and signaling pathways potentially involved in AD pathogenesis, supporting the framework's biological interpretability. Discussion: MOMHCA-SG effectively preserves feature integrity during high-dimensional processing, dynamically integrates heterogeneous omics data, and captures cross-modality relationships. These capabilities highlight its potential as a powerful tool for elucidating disease mechanisms and advancing precision medicine in neurodegenerative disorders.