A Pan-Cancer Benchmark of Clustering Strategies for Patient-Specific Gene Regulatory Networks
Cancer patients with clinically similar tumors can exhibit substantial differences in disease progression and treatment response. Patient-specific gene regulatory networks (GRNs) may capture regulatory heterogeneity that is not represented by gene expression alone. However, strategies for clustering patients using these high-dimensional networks have not been systematically evaluated and it remains unclear whether clustering based directly on millions of patient-specific regulatory edge weights is computationally feasible at scale. We benchmarked feature-selection and clustering approaches for patient-specific GRNs inferred with PANDA and LIONESS across 9,778 primary tumor samples from 33 cancer types. We compared three data representations: gene expression, gene indegrees summarizing regulatory input to each gene, and complete networks of transcription factor-gene edge weights. Twenty combinations of four feature-selection methods and five clustering algorithms were evaluated using consensus clustering. Performance was assessed based on mean silhouette score, PAC score, concordance, computational time, and memory usage. Standard deviation-based feature selection combined with k-means clustering achieved the strongest overall performance for both GRN-derived representations, with overall scores of 0.997 for edge weights and 0.985 for indegrees, while remaining computationally efficient. Edge-weight- and indegree-derived clusters showed clinical, biological, and prognostic associations that were complementary to those obtained from gene expression. Together, these findings provide a practical framework for network-based cancer patient stratification and highlight the potential of GRN-derived features to reveal clinically relevant tumor heterogeneity beyond gene expression alone.