Rapidly identifying protein complexes is significant to elucidate the mechanisms of macromolecular interactions and to further investigate the overlapping clinical manifestations of diseases. To date, existing computational methods majorly focus on developing unsupervised graph clustering algorithms, sometimes in combination with prior biological insights, to detect protein complexes from protein-protein interaction (PPI) networks. However, the outputs of these methods are potentially structural or functional modules within PPI networks. These modules do not necessarily correspond to the actual protein complexes that are formed via spatiotemporal aggregation of subunits. In this study, we propose a computational framework that combines supervised learning and dense subgraphs discovery to predict protein complexes. The proposed framework consists of two steps. The first step reconstructs genome-scale protein co-complex networks via training a supervised learning model of l₂-regularized logistic regression on experimentally derived co-complexed protein pairs; and the second step infers hierarchical and balanced clusters as complexes from the co-complex networks via effective but computationally intensive k-clique graph clustering method or efficient maximum modularity clustering (MMC) algorithm. Empirical studies of cross validation and independent test show that both steps achieve encouraging performance. The proposed framework is fundamentally novel and excels over existing methods in that the complexes inferred from protein cocomplex networks are more biologically relevant than those inferred from PPI networks, providing a new avenue for identifying novel protein complexes.