Abstract
In the general framework of the weighted gene co-expression network analysis (WGCNA), a hierarchical clustering algorithm is commonly used for module definition. However, hierarchical clustering depends strongly on the topological overlap measure (TOM), potentially assigning genes with similar expression patterns but low topological overlap to different modules. We propose a novel gene module clustering network (gmcNet) that simultaneously addresses single-level expression and TOM. gmcNet includes a “co-expression pattern recognizer" (CEPR) and a “module classifier." The CEPR incorporates expression features of single genes into the topological features of co-expressed ones. Validated using 4,976 genes from 20 native Korean cattle, gmcNet achieved superior performance in terms of modularity (0.261) and differentially expressed (DEM) signal (27.739) compared to traditional methods. Furthermore, gmcNet identified biological functionalities related to carcass weight, backfat thickness, intramuscular fat, and beef tenderness.
Introduction
Weighted Gene Co-expression Network Analysis (WGCNA) is a standard bioinformatics tool for identifying "modules"—clusters of genes that co-function to influence complex traits. Traditionally, WGCNA relies on Hierarchical Clustering (HC) based on a Topological Overlap Measure (TOM). However, HC suffers from information loss when raw expression data is compressed into a topological matrix and rigidity due to its "greedy" nature, where early assignment decisions cannot be reversed. This often leads to biologically "noisy" modules that fail to capture true functional relationships.
Methods and Materials
The proposed gmcNet framework treats gene clustering as an unsupervised learning task on a graph. It consists of two primary components:
- Co-expression Pattern Recognizer (CEPR): Utilizing a message-passing mechanism, the CEPR aggregates a gene’s individual expression features with those of its neighbors, using the TOM matrix as adjacency weights. This allows the model to learn from both individual expression levels and the strength of gene-gene connections.
- Module Classifier: A multi-layer perceptron (MLP) that maps CEPR-embedded features to module assignment probabilities.
Training is performed using the MinCutPool loss function, which optimizes for strong intra-module connectivity and distinct module separation. The model was evaluated using RNA-seq data from the longissimus dorsi muscle of 20 Korean native cattle (Hanwoo), focusing on 4,976 filtered genes.
Results and Discussion
A key finding of this study is that feature fusion is essential for robust module identification. As shown in the PCA analysis, neither raw expression nor network topology alone provides the clear separation achieved by the CEPR embedding. This suggests that the GNN successfully "denoises" the biological signal by integrating multiple data levels.
Quantitatively, gmcNet outperformed traditional methods across key metrics. It achieved a modularity score ($Q$) of 0.261, significantly higher than HC (0.219) or K-means (0.138). Additionally, the "DEM signal," which measures the correlation between modules and external traits, reached 27.739 for gmcNet, compared to 18.618 for traditional HC.
The model identified 11 significant module-trait interactions. Notably, it linked immune system pathways to beef tenderness (WBSF), a novel finding in Hanwoo cattle. Module K1 was enriched for metabolic processes related to carcass weight, while K7 emerged as a critical module for tenderness.
Conclusion
gmcNet provides a powerful alternative to hierarchical clustering in WGCNA by leveraging the strengths of Graph Neural Networks. While the current implementation requires a pre-specified number of clusters ($k$) and faces scaling challenges for full transcriptomes, its ability to integrate raw expression with topology leads to more biologically meaningful insights. Future work will focus on automated $k$ determination and improved noise handling.
Summary updated