Empirical clinical studies on the human interactome and phenome not only illustrates prevalent phenotypic overlap and genetic overlap between diseases, but also reveals a modular organization of the genetic landscape of human disease, providing new opportunities to reduce the complexity in dissecting the phenotype-genotype association. We here introduce a network-module based method towards phenotype-genotype association inference and disease gene identification. This approach incorporates protein-protein interaction network, phenotype similarity network and known phenotype-genotype associations into an assembled network. We then decomposes the resulted network into modules (or communities) wherein we identified and prioritized the disease genes from the candidates within the loci associated with the query disease using a linear regression model and concordance score. For the known phenotype-gene associations in the OMIM database, we used the leave-one-out validation to evaluate the feasibility of our method, and successfully ranked known disease genes at top 1 in 887 out of 1807 cases. Moreover, applying this approach on 850 OMIM loci characterized by an unknown molecular basis, we propose high-probability candidates for 81 genetic diseases.