From the drug discovery perspective, combination therapy is recommended for cancer treatment due to its efficiency and safety compared to the common cytotoxic and single-targeted monotherapies. However, identifying effective drug combinations is time-and cost-consuming. Here, we offer a novel strategy for predicting potential drug combinations and patient subclasses by constructing multipartite networks using drug-response data on patient samples. In this study, we used Beat AML and GDSC, two comprehensive datasets based on patient-derived and cell line-based samples, to show the potential of multipartite network modelling in combinatorial cancer therapy. We used the median values of cell viability to compare drug potency and reconstruct a weighted bipartite network that models the interaction of drugs and biological samples. Then, clusters of network communities were identified in two projected networks based on the topological structure of the networks. Chemical structures, drug-target networks, protein– protein interactions, and signalling networks were used to corroborate the intra-cluster homogeneity. We further leveraged the community structures within the drug-based multipartite networks to discover effective multi-targeted drug combinations and synergy levels, which were supported with more evidence using the DrugComb and ALMANAC databases. Furthermore, we confirmed the potency of selective combinations of drugs against monotherapy in in vitro experiments using three acute myeloid leukaemia (AML) cell lines. Taken together, this study presents an innovative data-driven strategy based on multipartite networks to suggest potential drug combinations to improve AML treatment.