High-throughput mass-spectrometry technology has enabled genome-scale discovery of protein-protein interactions. Yet, computational inference of protein interaction networks and their functional modules from large-scale pull-down data is challenging. Over-expressed or "sticky" bait is not specific; it generates numerous false positives. This "curse" of the technique is also its "blessing" - the sticky bait can pull-down interacting components of other complexes, thus increase sensitivity. Finding optimal trade-offs between coverage and accuracy requires tuning multiple "knobs," i.e., method parameters. Each selection leads to a putative network, where each network in the set of "perturbed" networks differs from the others by a few added or removed edges. Identification of functional modules in such networks is often based on graph-theoretical methods such as maximal clique enumeration. Due to the NP-hard nature of the latter, the number of tunings to explore is limited. This paper presents an efficient iterative framework for sensitive and specific detection of protein complexes from noisy protein interaction data.