Abstract Background: Non-small cell lung cancer (NSCLC) is the most common subtype of lung cancer, characterizing ~80% of lung cancer diagnoses. Thus, the development of novel therapies for NSCLC is of primary importance in oncology. The aim of this study was to develop a novel cancer systems biology approach to understand the complex interactions in NSCLC and predict the response to target treatments. We combined proteomic experimental data with a machine learning tool to discover the diverse signaling networks of both KRAS mutant (MUT) and wild-type (WT) NSCLC cell lines (CL). Material and Methods: Eight NSCLC CL (5 KRAS MUT and 3 KRAS WT) were subjected to reverse phase protein microarray (RPPA) to explore the activation level of 183 proteins treated with selumetinib (SE) and its combination with everolimus (EV) or tamoxifen (TA). Measures were taken at 6 time points: 5 minutes (min), 1, 2 , 6 and 24 hours (h). RPPA was performed also on the baseline and on CL with only dimethyl sulfoxide. Recursive Feature Elimination with Support Vector Machine (RFE-SVM) was used to identify the subset of proteins/features that optimally separated samples in 2 groups. In order to build the interaction network from this subset and browse the main pathways involved, we used Reactome FI and the Pathway Enrichment tool, two Cytoscape plugins for network-based data analysis. Results: We applied RFE-SVM to discover the most divergent proteins between control and treated samples in the whole CL panel. For all the 3 drug combinations, the network in output contained not only the target of the corresponding drug, but also showed the activation of other parallel pathways. The analysis revealed also that all drugs were mainly effective in a short time, since the number of significant proteins was considerably higher at 5 min and 1 h. For instance, SE treatment activated NGFR, mTOR, PI3K-Akt and proteoglycans pathways in addition to the MAPK pathway. Network at 1 h contained 25 proteins linked by 76 interactions, while at 6 and 24 h the distinctive subsets of features included only 2 and 4 proteins. Then, we applied RFE-SVM to analyze KRAS MUT vs WT CLs. The algorithm showed that treated KRAS MUT CL had a more complex signaling network than the WT ones. Indeed, in SE WT samples MER and ERK were the main relevant proteins to distinguish from the control. In treated KRAS MUT samples, a complex network of 70 proteins and 278 interactions was active during the first hour. Conclusion: The proposed framework processes high-throughput data and identifies new possible molecular targets in cancer research. This approach is highly flexible for a multitude of data and purposes. We applied it to a RPPA dataset of NSCLC CL and it revealed that all drug combinations had fast dynamics and that KRAS MUT CL had a more complicated response due to a higher number of active pathways. Citation Format: Chiara Antonini, Lorenzo Tomassoni, Elisa Baldelli, Vienna Ludovini, Sara Baglivo, Mariaelena Pierobon, Emanuel F Petricoin, Lucio Crinò, Paolo Valigi, Fortunato Bianconi. Dynamic network topologies analysis in proteomics data of NSCLC cell lines using a new pipeline based on machine learning tools [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 2254.