Distributed control offers practical trade-off between globally (plant-wide) valid, computationally expensive centralized and locally effective decentralized architectures. While the development of distributed control theory (algorithms, stability, convergence properties, etc.) has made a significant progress in the last decade, decomposition of a system into an optimal distributed structure (classification of controller variables) has received considerably less attention. This paper focuses on achieving such a decomposition for integrated process systems with the objective of maximizing the strength of input-output impact and thus reducing the strength of inter-controller interactions. A graph-theoretic framework is used to abstract input-state-output dependence of a dynamical system and control-relevant interactions are quantified using weighted digraphs. It is shown that community detection in this weighted digraph is equivalent to distributed structure corresponding to strong input-output impact (quantified through the corresponding gain). Two relevant example systems are considered to illustrate the proposed framework and it is shown that the optimal distributed architectures adhere to the underlying structure of strong and weak inter-variable interactions. (C) 2019 Elsevier Ltd. All rights reserved.