Network Fragmentation Occurs Earlier in LUAD Tumor Gene Co-expression Networks Than in Normal Lung

Introduction

Cancer transcriptomes exhibit widespread changes in gene regulation and co-regulation. Gene co-expression networks represent genes as nodes and co-expression relationships as edges; their large-scale topology can be probed by progressively increasing a correlation threshold, effectively removing weaker edges. The resulting decline of the giant connected component (GCC) provides a direct, interpretable measure of network-level robustness.

Here we test whether LUAD (lung adenocarcinoma) tumor networks show reduced structural stability relative to normal lung tissue. We quantify fragmentation across thresholds using multiple complementary definitions of the critical threshold r* and validate robustness via tumor-only percolation, gene-set size sensitivity analysis, and bootstrap resampling.

Methods

Data and preprocessing

LUAD primary tumor RNA-seq expression (n=513) and normal lung expression (n=578) were analyzed using TPM values with log transformation (X = log(1+TPM)). Genes with zero variance were removed.

Matched (joint) gene set selection

For cross-cohort comparison, analyses were restricted to a matched node set: the top 4,000 genes ranked by combined variance across tumor and normal cohorts (computed on the jointly filtered gene universe). This “joint set” design ensures apples-to-apples comparison of network topology.

Network construction

Pairwise Pearson correlations were computed between genes across samples. Only positive correlations were retained. For a threshold r, an undirected edge was present if C_ij ≥ r.

Connectivity and critical threshold definitions

For each r, GCC(r) was computed as the fraction of nodes in the largest connected component. The critical threshold r* was defined using three criteria: (i) first crossing of GCC<0.9, (ii) first crossing of GCC<0.5, and (iii) the threshold of maximal negative slope (knee proxy).

Tumor-only percolation and gene-set size robustness

To characterize intrinsic LUAD behavior, tumor-only percolation was evaluated using the top 8,000 tumor genes by variance. Gene-set size robustness was assessed by repeating tumor-only analysis for 4,000–16,000 genes.

Bootstrap validation

Bootstrap resampling (60 replicates) randomly sampled 2,000 genes without replacement per replicate from the joint gene pool. For each replicate, r* values were recomputed for tumor and normal networks.

Results

Tumor networks fragment earlier than normal lung (matched joint gene set)

Using the joint top-4,000 gene set, LUAD tumor networks lost global connectivity at lower thresholds than normal lung. The first crossing of GCC below 0.9 occurred at tumor r* = 0.420 ± 0.006 versus normal r* = 0.584 ± 0.006 (Δ ≈ 0.164). Under GCC<0.5, tumor r* = 0.634 ± 0.005 versus normal r* = 0.730 ± 0.002 (Δ ≈ 0.096). The knee proxy (maximum slope) remained lower in tumors (0.732 ± 0.031) than normal lung (0.767 ± 0.013; Δ ≈ 0.035).