With the advancements of high-throughput sequencing technology, several recent studies addressed the clinical/phenotypic stratification of samples by utilizing transcriptome data. However, existing stratification methods lack efficient utilization of gene interaction information, and furthermore, handling more than 20,000 genes causes the curse of high dimensionality that hinders elucidating the linkage between genetic profiles and clinical/phenotypic differences. To overcome these challenges, we propose a network-based two-step computational framework. We first reduce dimensions of transcriptome to a few tens of dimensions by mapping transcriptome to protein interaction network followed by performing network propagation algorithm and clustering analysis. Then, each network is converted into a single numeric metric by utilizing information theoretic quantification of gene expression abnormality, which results in a single sample mapping to a metric space generated by each subnetwork in the form of vectors. The proposed network-based stratification method was used to analyses Pan-Caner dataset and Oryza sativa dataset. Extensive experiments showed that our method generates a metric space that captures data-specific biological functions and improves the stratification performance compared to existing methods. Therefore, the proposed method successfully stratified the samples, addressing the problem in the complex gene space. The proposed method is implemented in Python and available at https://github.com/Sunginyoung/net_stratification.