Abstract
Drug combinations have emerged as a promising therapeutic approach in cancer treatment, aimed at overcoming drug resistance and improving the efficacy of monotherapy regimens. However, identifying effective drug combinations has traditionally been time-consuming and often dependent on chance discoveries. Therefore, there is an urgent need to explore alternative strategies to support experimental research. In this study, we propose network-based prediction models to identify potential drug combinations for 11 types of cancer. Our approach involves extracting 55,299 associations from literature and constructing human protein interactomes for each cancer type. To predict drug combinations, we measure the proximity of drug-drug relationships within the network and employ a correlation clustering framework to detect functional communities. Finally, we identify 61,754 drug combinations. Furthermore, we analyze the network configurations specific to different cancer types and identify 30 key genes and 21 pathways. The performance of these models is subsequently assessed through in vitro assays, which exhibit a significant level of agreement. These findings represent a valuable contribution to the development of network-based drug combination design strategies, presenting potential solutions to overcome drug resistance and enhance cancer treatment outcomes.