Abstract
Identifying functional gene modules is essential for understanding the molecular mechanisms of complex diseases like hepatocellular carcinoma (HCC). This paper presents GCNA-Kpca, a novel hybrid algorithm that integrates network topology from the Newman algorithm with iterative refinement via a modified K-means framework. By optimizing modularity and allowing for iterative cluster reassignment, GCNA-Kpca overcomes the limitations of traditional hierarchical methods like WGCNA. Our results demonstrate that GCNA-Kpca achieves significantly lower error rates and higher biological significance. When applied to HCC data, the algorithm identified 10 key biomarkers, including NOP56 and NOP58, which are reported here for the first time in the context of liver cancer.
Introduction
In systems biology, genes operate within coordinated modules rather than in isolation. Accurately identifying these modules is a prerequisite for discovering reliable biomarkers and therapeutic targets. While Weighted Gene Co-expression Network Analysis (WGCNA) is widely used, it suffers from two major flaws: a lack of explicit modularity optimization and the "irreversibility" of hierarchical clustering, where initial gene assignments cannot be corrected even if they fit better elsewhere after further analysis.
Methods: The GCNA-Kpca Algorithm
The proposed GCNA-Kpca algorithm bridges network science and machine learning through a three-stage workflow:
- Network Construction: A Gene Co-expression Network (GCN) is built using Pearson correlation coefficients.
- Initial Partitioning: The Newman algorithm—a community detection method—identifies initial modules based on network topology, determining the optimal number of clusters ($K$) and their initial centers.
- Iterative Refinement: A modified K-means framework refines these clusters. Unlike standard K-means, this approach uses a correlation-based distance metric ($1 - |cor(g, C)|$) and employs Principal Component Analysis (PCA) to calculate "Module Eigengenes" as cluster centers, ensuring they represent the dominant expression pattern.
Results and Discussion
The performance of GCNA-Kpca was evaluated against seven other algorithms using HCC expression profiles from TCGA and GEO datasets.
Comparative Performance
GCNA-Kpca significantly outperformed existing methods across all metrics. It achieved an error rate of 0.06, a substantial improvement over WGCNA (0.29) and the standalone Newman algorithm (0.25). Furthermore, as shown in Figure 3, the biological significance score (Sig) for GCNA-Kpca reached 956.52, nearly double that of WGCNA (499.40), indicating that the identified modules are more functionally coherent.
Identification of Key Biomarkers
The algorithm identified 10 key genes in HCC: CDC20, CCNB1, EIF4A3, H2AFX, NOP56, RFC4, NOP58, AURKA, PCNA, and FEN1. Notably, NOP56 and NOP58—nucleolar proteins involved in ribosome biogenesis—were identified as novel HCC biomarkers. Survival analysis confirmed that these 10 genes possess high prognostic significance, suggesting their potential as clinical biomarkers for liver cancer.
Conclusion
The GCNA-Kpca algorithm provides a robust framework for gene module identification by combining the structural insights of community detection with the iterative precision of K-means. Its success in identifying novel biomarkers for HCC highlights its potential for broader application in cancer research. Future work will focus on experimental validation of NOP56 and NOP58 and extending the algorithm to multi-omic data.
Summary updated