Adaptive NetworkProfiler for Identifying Cancer Characteristic-Specific Gene Regulatory Networks

There is currently much discussion about sample (patient)-specific gene regulatory network identification, since the efficiently constructed sample-specific gene networks lead to effective personalized cancer therapy. Although statistical approaches have been proposed for inferring gene regulatory networks, the methods cannot reveal sample-specific characteristics because the existing methods, such as an L₁-type regularization, provide averaged results for all samples. Thus, we cannot reveal sample-specific characteristics in transcriptional regulatory networks. To settle on this issue, the NetworkProfiler was proposed based on the kernel-based L₁-type regularization. The NetworkProfiler imposes a weight on each sample based on the Gaussian kernal function for controlling effect of samples on modeling a target sample, where the amount of weight depends on similarity of cancer characteristics between samples. The method, however, cannot perform gene regulatory network identification well for a target sample in a sparse region (i.e., for a target sample, there are only a few samples having a similar characteristic of the target sample, where the characteristic is considered as a modulator in sample-specific gene network construction), since a constant bandwidth in the Gaussian kernel function cannot effectively group samples for modeling a target sample in sparse region. The cancer characteristics, such as an anti-cancer drug sensitivity, are usually nonuniformly distributed, and thus modeling for samples in a sparse region is also a crucial issue. We propose a novel kernel-based L₁-type regularization method based on a modified k-nearest neighbor (KNN)-Gaussian kernel function, called an adaptive NetworkProfiler. By using the modified KNN-Gaussian kernel function, our method provides robust results against the distribution of modulators, and properly groups samples according to a cancer characteristic for sample-specific analysis. Furthermore, we propose a sample-specific generalized cross-validation for choosing the sample-specific tuning parameters in the kernel-based L₁-type regularization method. Numerical studies demonstrate that the proposed adaptive NetworkProfiler effectively performs sample-specific gene network construction. We apply the proposed statistical strategy to the publicly available Sanger Genomic data analysis, and extract anti-cancer drug sensitivity-specific gene regulatory networks.