Perturbation-driven sensitive gene discovery in colorectal cancer.

Colorectal cancer (CRC) progression involves complex gene regulatory interactions, yet conventional analyses often overlook genes that, while not canonical drivers, may contribute to network-level instability. In this study, we propose a dynamic computational framework that integrates time-series co-expression networks with an autoregressive neural network and local network entropy (ARNN-LNE) to quantify perturbation-induced changes in gene regulatory stability. Rather than relying on static correlation patterns, this framework evaluates how in silico perturbations influence network entropy to identify network-sensitive genes.Applying this approach to CRC datasets, we identify a set of candidate sensitive genes, including MATCAP1, FAM107B, SNX24, and SLC26A2 in human, and mt-Co1 in mouse, which exhibit pronounced entropy responses under perturbation conditions. These genes are not readily captured by conventional differential expression analysis, suggesting complementary information from a network dynamics perspective. Cross-species analysis further indicates partial consistency in identified sensitive genes across human and mouse datasets, supporting the robustness of the proposed framework.Furthermore, expression-based classification analysis suggests that these genes exhibit moderate discriminative ability for distinguishing disease states (ROC-AUC ≈ 0.86; PR-AUC ≈ 0.76), indicating potential utility for further investigation. Overall, this study presents perturbation-entropy profiling as a computational framework for identifying network-sensitive genes, providing a hypothesis-generating approach for exploring gene regulatory dynamics in cancer systems biology.