Generalized information criteria for personalized gene network inference.

Identifying individual genomic characteristics is a critical focus in personalized therapies. To reveal targets in such therapies, we considered personalized gene network analysis using kernel-based $L_1$ -type regularization methods. In kernel-based $L_1$ -type regularized modeling, selecting optimal regularization parameters is crucial because edge selection and weight estimation depend heavily on such parameters. Furthermore, selecting a kernel bandwidth that controls sample weighting is vital for personalized modeling. Although cross-validation and information criteria (i.e., AIC and BIC) are often used for parameter selection, such traditional techniques are computationally expensive or unsuitable for approaches based on estimation techniques other than maximum likelihood estimation. To overcome these issues, we introduced a novel evaluation criterion in line with the generalized information criterion (GIC), which relaxes the assumption of maximum likelihood estimation, making it suitable for personalized gene network analysis based on various estimation techniques. Monte Carlo simulations demonstrated that the proposed GIC outperforms existing evaluation criteria in terms of edge selection and weight estimation. Acute myeloid leukemia (AML) drug sensitivity-specific gene network analysis revealed critical molecular interactions to uncover ALM drugs resistant mechanism. Notably, PIK3CD activation and RARA/RELA suppression are crucial markers for improving AML chemotherapy efficacy. We also applied our strategy for gastric cancer drug sensitivity analysis and uncovered personalized therapeutic targets. We expect that the proposed sample specific GIC will be a useful tool for evaluating personalized modeling, including in sample characteristic-specific gene networks analysis.