In the past decade, a large number of genetic biomarkers have been discovered through large-scale genome wide association studies (GWASs) in Alzheimer's disease (AD), such as APOE , TOMM40 and CLU . Despite this significant progress, existing genetic findings are largely passengers not directly involved in the driver events, which presents challenges for replication and translation into targetable mechanisms. In this paper, leveraging the protein interaction network, we proposed a modularity-constrained Lasso model to jointly analyze the genotype, gene expression and protein expression data. With a prior network capturing the functional relationship between SNPs, genes and proteins, the newly introduced penalty term maximizes the global modularity of the subnetwork involving selected markers and encourages the selection of multi-omic markers with dense functional connectivity, instead of individual markers. We applied this new model to the real data in ROS/MAP cohort for discovery of biomarkers related to cognitive performance. A functionally connected subnetwork involving 276 multi-omic biomarkers, including SNPs, genes and proteins, were identified to bear predictive power. Within this subnetwork, multiple trans-omic paths from SNPs to genes and then proteins were observed, suggesting that cognitive performance can be potentially affected by the genetic mutations due to their cascade effect on the expression of downstream genes and proteins.