Network-Based Dynamical Simulation of Genomic Co-expression Profiles Predicts Synergistic Target Combinations in Lung and Breast Cancers

Tumor heterogeneity and compensatory pathway activation present major obstacles to targeted monotherapies in oncology. We developed a network-based dynamical simulation model to evaluate combination therapies in silico on 515 lung adenocarcinoma (LUAD) and 1,097 breast cancer patient samples from the TCGA breast invasive carcinoma cohort (designated as BRCA; throughout this work, BRCA refers to the TCGA disease acronym and not the BRCA1/BRCA2 breast cancer susceptibility genes), mapping the co-expression topology of a genome-informed 500-gene network. Using a GPU-accelerated vectorized PyTorch implementation, we screened all 124,750 possible dual-target combinations per cohort in a single parallel batch (~7 seconds per cohort), and subsequently validated the temporal dynamics of the top-ranked pairs through an optimized batch 3D tensor simulation on GPU. The model identified MTOR-centric combinations as the top synergistic targets in LUAD: MTOR + UGT8 (Rank 1), FUT9 + MTOR (Rank 2), and the biologically notable GRIA2 + MTOR (Rank 4). In BRCA, CEACAM6-centric combinations dominated---CEACAM6 + CST4 and CEACAM6 + PEBP4---with LRRC31 + SHISA2 emerging as a novel stemness-disruption candidate. COL11A1 was the most overexpressed gene in LUAD within the 500-gene panel; MMP13 led in BRCA. Projections of these dynamic states into 3D state-space trajectories and temporal evolution profiles demonstrate how multi-target interventions steer the pathological attractor states toward network-wide stress reduction. Notably, the top synergistic targets identified by the model---most critically MTOR in LUAD---do not correspond to the highest-centrality nodes in the static co-expression network (MTOR centrality: 1.30 vs. network maximum: 28.17), demonstrating that dynamical regulatory impact and static correlation degree are dissociated properties. These results demonstrate the capacity of dynamical co-expression network models to recover tissue-specific therapeutic vulnerabilities from observational genomic data at genome-informed scale.