Through the analysis of forward and feedback links of a biochemical network based on the input-output polarity and the principle of optimal usage of resources, we are able to decompose a large network into modules. For signal transduction network,  a hierarchical structure of feedbacks is identified with local feedbacks exerting quick strong control and the global ones adjusting the dynamics in a slow and subtle way. Moreover, at the fork of forward pathways or the  intersection of the forward and  feedback edges, signaling hubs are often detected which constitute bottlenecks of the whole network.