Summary Alzheimer’s disease (AD) is the most common form of neurodegenerative dementia and has an astounding impact at individual and societal levels. As the early-stage cognitive degeneration, mild cognitive impairment (MCI) has a high chance to convert to AD. Effective and early prediction of such conversion is of great importance for risk stratification, patient management, and possible symptomatic treatments. Identification of an MCI-AD conversion end point is also important for clinical trials for better evaluating the effectiveness of therapeutic interventions. Recent studies have shown that multi-modalities neuroimaging can offer a more comprehensive characterization for the MCI-AD conversion, revealing the physiologic underpinning of the clinical states, and ultimately result in higher prediction accuracy based on the multi-modal imaging biomarker. Advancement in deep learning, especially deep Graph Convolutional Networks, has provided us with powerful tools in modeling the multi-modal neuroimaging data on the brain networks. However, despite the high prediction accuracy of AD in literature, multi- modal imaging diagnostic still lacks generalizability and robustness in dealing with data from other sites/populations due to the combined effect of relatively smaller sample sizes and potential bias in the sample labels. In this proposal, we will investigate the interaction among structural, functional, and proteinopathies networks in MCI and AD patients via a contrastive learning-based, multigraph representation framework on the multi-modal neuroimaging data of MRI, fMRI and PET modalities. The proposed framework will be used to identify and evaluate a multi-modal image biomarker for the AD conversion in MCI population from a multi-site dataset. By analyzing the spatial and populational patterns of the identified multi-modal image biomarker, we will be able to discover novel neuroscientific and biological mechanisms of the MCI-AD conversion.