Sample-Specific Generalized Cross-Validation for Gene Network Analysis of Cytarabine Response in Cancer Cell Lines

Sample-specific gene regulatory network analysis can reveal molecular heterogeneity associated with individual characteristics, such as anticancer drug sensitivity. The varying coefficient model with kernel-based L1 regularization enables the estimation of such networks, but its performance depends strongly on hyperparameter selection. Conventional cross-validation is computationally intensive and provides only an averaged evaluation across samples, limiting its suitability for sample-specific analysis. To address these limitations, we propose doubleS-GCV, a sample-specific generalized cross-validation criterion for selecting hyperparameters in sample-specific gene network estimation. DoubleS-GCV provides a separate model evaluation for each sample while substantially reducing computational burden. Monte Carlo simulations demonstrated that doubleS-GCV achieved accurate gene selection and network estimation and outperformed conventional information criteria, including AIC, BIC, AICC, and HQC. Application to GDSC cancer cell lines identified Cytarabine sensitivity-specific gene networks and candidate biomarkers supported by previous studies. The estimated networks also exhibited nonlinear structural changes across Cytarabine sensitivity levels, indicating that molecular interactions vary with drug response. These results demonstrate that doubleS-GCV provides an efficient and reliable model selection framework for sample-specific gene network analysis.