Patient-specific gene co-expression networks reveal novel subtypes and predictive biomarkers in lung adenocarcinoma.

Lung adenocarcinoma (LUAD) is a highly heterogenous and aggressive form of non-small cell lung cancer (NSCLC). The use of genome-wide gene co-expression networks (GCNs) has been paramount to describe changes in the transcriptional regulatory programs found between diseased and healthy states of LUAD. Recently, studies have shown that multiple cancerous phenotypes share a distinct GCN architecture, suggesting that network topology holds promise for understanding disease pathology. However, conventional GCN inference methods struggle to capture the inherent context-specificity within a patient population, thus flattening its heterogeneity. To address this issue, the use of single-sample network (SSN) modelling has emerged as a promising solution into studying heterogeneous traits of cancer through network-based approaches. Here, we reconstructed patient-specific GCNs (n=334) using the LIONESS equation and mutual information as the network inference method. Unsupervised analysis revealed six novel LUAD subtypes based on inter-patient network similarity, each with distinct network motifs reflecting unique biological programs. Supervised analysis, employing regularized Cox regression, identified 12 genes (CHRDL2, SPP2, VAC14, IRF5, GUCY1B1, NCS1, RRM2B, EIF5A2, CCDC62, CTCFL, XG, and TP53INP2) whose weighted degree in SSNs is predictive of patient survival in LUAD. These findings suggest that topological features of SSNs offer valuable insights into the context-specific nature of LUAD malignancy, highlighting the potential of SSN-based approaches for further research.