Statistical Inference of Phenotype-Specific Molecular Mechanisms from Cell Line-Specific Gene Regulatory Networks...

Gene regulatory networks differ substantially across individual cell lines, and population-level network inferences often fail to capture the underlying biological heterogeneity. To better capture this heterogeneity, cell line-specific gene network analysis is required. However, interpreting such high-dimensional cell line-specific networks remains a major challenge in the field of network biology. One interpretative approach is to identify differentially regulated gene networks (DGNs) between phenotypes because these networks can highlight phenotype-specific regulatory mechanisms. Although several methods have been proposed for DGN analysis, they are not suitable for cell line-specific gene network analysis, which is characterized by pronounced heterogeneity across individual networks. To address this problem, we proposed a novel statistical method for identifying DGNs in a cell line-specific scenario. The proposed framework integrates cell line-specific network estimation, a Kullback-Leibler divergence-based comparison of multivariate distributions, and a DKL-ratio statistic to quantify between-phenotype heterogeneity relative to within-phenotype homogeneity. Our method evaluates both between-phenotype heterogeneity and within-phenotype homogeneity, ensuring the robust detection of phenotype-specific network structures. Through Monte Carlo simulation studies, we systematically evaluated the performance of the proposed method and demonstrated that our strategy consistently outperformed existing methods in terms of accuracy, precision, true positive rate (TPR), true negative rate (TNR), and F-measure across diverse network structures and mean shift scenarios. Statistical significance was assessed using a permutation-based framework, confirming that the identified networks are unlikely to arise from random variation. We further applied our strategy to Quizartinib sensitivity-specific gene network analysis and identified immune-related subnetworks enriched in antigen processing and presentation pathways. These subnetworks included hub genes such as IFIT1, PSMB9, and HLA-B, which are known to be associated with immune evasion and drug resistance in acute myeloid leukemia. Our findings demonstrate that the proposed method enables statistically reliable and biologically interpretable identification of phenotype-specific gene regulatory mechanisms, providing insights into potential therapeutic targets.