Diagnosis-guided multimodal adversarial graph masked autoencoder network for association analysis between brain imaging and gene expression in Alzheimer's disease.
Authors
Affiliations (3)
Affiliations (3)
- MOE Key Laboratory of Information Fusion Technology, School of Automation, Northwestern Polytechnical University, Xian 710072, China. Electronic address: [email protected].
- MOE Key Laboratory of Information Fusion Technology, School of Automation, Northwestern Polytechnical University, Xian 710072, China. Electronic address: [email protected].
- School of Future Technology, South China University of Technology, Guangzhou 511442, China. Electronic address: [email protected].
Abstract
Identifying the associations between brain imaging and gene expression data is crucial for uncovering potential biomarkers of Alzheimer's disease (AD). Although various methods have been developed for association analysis using imaging genetics data, most rely on single nucleotide polymorphisms (SNPs) that reflect genetic variation, while overlooking gene expression data that reflect the impact of environmental factors on the progression of AD. Moreover, existing imaging-genetics association methods rely on unsupervised learning, thus fail to utilize the diagnostic information related to AD. To address these limitations, we propose a novel diagnosis-guided Multimodal Adversarial Graph Masked Autoencoder network (MAGMA), an end-to-end deep learning framework to integrate functional magnetic resonance imaging (fMRI) and gene expression data for association analysis between brain imaging and gene expression in AD. MAGMA comprises two major components: (1) a multimodal graph representation module, which is designed to extract the diagnosis-guided latent embeddings of brain fMRI and gene expression data by employing two Adversarial Graph Masked Autoencoders (AGMAs) with random feature masking, and the diagnosis information is incorporated to guide the learning of disease-relevant representations; and (2) An association analysis module, which models nonlinear relationships between brain imaging and gene expression data by nonlinearly mapping gene representations to imaging representations. To demonstrate the validity of MAGMA, we conducted experiments on Alzheimer's Disease Neuroimaging Initiative dataset and analyzed the results from diverse perspectives. Experimental results demonstrate that MAGMA can identify potential disease-related biomarkers, many of which are biologically interpretable and supported by existing AD-related studies. These findings suggest that MAGMA may offer new insights into the pathological mechanisms of AD and contribute to its early diagnosis.