Weighted Gene Co-expression Network Analysis (WGCNA) is widely used for identifying co-expression modules in transcriptomic studies due to its computational efficiency. However, its reliance on linear models may limit sensitivity to complex, nonlinear gene interactions, particularly in datasets with low variability like sepsis and septic shock, where shared pathogenesis results in subtle transcriptional differences. To overcome these limitations, we introduce a novel approach called Nonlinear Overlapping Variability Analysis of Gene Networks or NOVAGene which directly contrasts the individual nonlinear gene co-expression networks between conditions, effectively mitigating the impact of sample heterogeneity. Using a local sepsis dataset, we evaluated WGCNA’s ability to capture co-expression modules across varying biological variability. While effective in highly variable datasets, such as those comparing diseased and healthy individuals, WGCNA struggled to differentiate the transcriptome from patients who progressed to septic shock from those who did not due to its reliance on linear approximations and sample pooling. In contrast, NOVAGene identifies biologically meaningful modules across both low and high variable datasets. It effectively distinguishes patients who will develop septic shock from those who will not, revealing shared core immune processes but distinct regulatory dynamics that differentiate these two clinical outcomes. Our enhancement to gene co-expression analysis using NOVAGene offers a more refined approach to sepsis prognostication, providing insights into the subset of biological processes that are innately dysregulated in patients who are susceptible to septic shock. NOVAGene may facilitate the detection of subtle variations in key biomarkers that could improve the prediction of disease outcomes. Despite its higher computational demands and potential statistical limitations, NOVAGene effectively captures nonlinear transcriptional patterns, enhancing the resolution of disease network analysis in biologically similar datasets with minimal transcriptional variation.
Bacong et al. (Tue,) studied this question.