A protein network shows physical interactions as well as functional associations. An important
usage of such networks is to discover unknown members of partially known complexes and
pathways. A number of methods exist for such analyses, and they can be divided into two main
categories based on their treatment of highly connected proteins. In this paper, we show that
methods that are not affected by the degree (number of linkages) of a protein give more accurate
predictions for certain complexes and pathways. We propose a network flow-based technique
to compute the association probability of a pair of proteins. We extend the proposed technique
using hierarchical clustering in order to scale well with the size of proteome. We also show that
top-k queries are not suitable for a large number of cases, and threshold queries are more meaningful
in these cases. Network flow technique with clustering is able to optimize meaningful
threshold queries and answer them with high efficiency compared to a similar method that uses
Monte Carlo simulation.