Accurate diagnose of Alzheimer’s disease (AD), especially in its early stage, is very important for the possible delay and early treatment of the disease. Many researches indicate that the change of the expression level of plasma proteins will lead to the emergence of AD, which creates a new focus on finding essential proteins that affects AD. Nearly all the existing models were constructed on the independent hypothesis, where each candidate protein is treated independently. However, proteomics studies suggest that the essential proteins in cells do not act independently under illness conditions. Instead, they perform biological functions in groups, where each group is composed of proteins with similar functions. Accordingly, we propose a manifold regularized sparse group-lasso (i.e., MSGL) method to identify essential proteins that will incur AD. Specifically, we firstly propose a novel method to calculate the similarity between pairs of proteins with the aid of Gene Ontology, and then use the spectral clustering algorithm to divide the proteins into groups. Next, we adopt a sparse group-lasso model to select proteins according to both intra-group sparsity and inter-group sparsity, which ensures only a small number of proteins in a few of groups will be selected. In addition, we also introduce a manifold-based Laplacian regularizer to preserve the data distribution information. Finally, support vector machine (SVM) is used to classify AD patients. The experimental results show that our method can achieve a classification accuracy of 97.5%, a sensitivity of 97.6%, and specificity of 97.4%, demonstrating great potential in classification of AD patients.