Use of a graph neural network to the weighted gene co-expression network analysis of Korean native cattle

Hyo-Jun Lee, Yoonji Chung, Ki Yong Chung, Young‐Kuk Kim, Jun Heon Lee, Yeong Jun Koh, Seung Hwan Lee

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. In other words, this algorithm may assign two genes with low topological overlap to different modules even though their expression patterns are similar. Here, a novel gene module clustering algorithm for WGCNA is proposed. We develop a gene module clustering network (gmcNet), which simultaneously addresses single-level expression and topological overlap measure. The proposed gmcNet includes a "co-expression pattern recognizer" (CEPR) and "module classifier". The CEPR incorporates expression features of single genes into the topological features of co-expressed ones. Given this CEPR-embedded feature, the module classifier computes module assignment probabilities. We validated gmcNet performance using 4,976 genes from 20 native Korean cattle. We observed that the CEPR generates more robust features than single-level expression or topological overlap measure. Given the CEPR-embedded feature, gmcNet achieved the best performance in terms of modularity (0.261) and the differentially expressed signal (27.739) compared with other clustering methods tested. Furthermore, gmcNet detected some interesting biological functionalities for carcass weight, backfat thickness, intramuscular fat, and beef tenderness of Korean native cattle. Therefore, gmcNet is a useful framework for WGCNA module clustering.

Introduction

In systems biology, Weighted Gene Co-expression Network Analysis (WGCNA) is the standard method for grouping thousands of genes into functional "modules." Traditionally, this process relies on Hierarchical Clustering (HC) based on a Topological Overlap Measure (TOM)—a metric reflecting how many neighbors two genes share. However, HC has significant limitations. First, converting raw expression data into a TOM matrix often discards subtle, single-level expression features, potentially splitting genes with similar expression patterns into different modules. Second, HC is a rigid, non-reversible algorithm; once a gene is assigned to a branch, the decision cannot be corrected, which is problematic for complex biological networks.

The gmcNet Framework

To address these challenges, researchers developed gmcNet (gene module clustering network), a Graph Neural Network (GNN) framework that learns module assignments by respecting both raw data and network structure. The architecture consists of two main components:

  1. Co-expression Pattern Recognizer (CEPR): This component uses Message Passing (MP) operations on the TOM. It integrates a gene’s own expression features with those of its neighbors, handling positive and negative co-expression separately to capture nuanced biological pathways.
  2. Module Classifier: A multi-layer perceptron (MLP) that uses CEPR-embedded features to calculate the probability of a gene belonging to a specific module via softmax activation.
PCA visualization showing that CEPR-embedded features provide superior cluster separation compared to raw expression or TOM alone.

The model is trained using a MinCutPool loss function, which encourages grouping strongly connected genes while ensuring modules are distinct and reasonably sized.

Performance and Biological Insights

The gmcNet model was validated using a dataset of 4,976 genes from 20 native Korean cattle (Hanwoo), focusing on economic traits such as carcass weight (CWT), backfat thickness (BF), intramuscular fat (IMF), and beef tenderness (WBSF).

Quantitative Superiority

Compared to traditional methods like HC, K-means, and K-medoids, gmcNet demonstrated superior performance across key metrics. It achieved a modularity score ($Q$) of 0.261 and a Differentially Expressed Module (DEM) signal of 27.739, significantly outperforming the traditional HC baseline.

Comparison of DEM signal and Modularity (Q) across different clustering methods and cluster numbers (k).

Biological Relevance

Beyond statistical performance, gmcNet identified critical biological links. It successfully pinpointed "hub genes" supported by literature, such as LAMTOR5 for growth and ACSL3 for fat metabolism. Notably, the model uncovered a novel link between immune system pathways and beef tenderness (WBSF). While immune responses are known to affect weight gain, their role in meat texture provides a potential new biomarker for cattle breeding.

Conclusion

gmcNet represents a significant advancement in transcriptomics by moving from rigid statistical recipes to flexible, learnable models. While future work aims to automate the selection of the optimal number of clusters ($k$) and improve noise handling, gmcNet currently provides a robust and biologically insightful framework for gene co-expression analysis.

Summary updated

Abstract

In the general framework of the weighted gene co-expression network analysis (WGCNA), a hierarchical clustering algorithm is commonly used to module definition. However, hierarchical clustering depends strongly on the topological overlap measure. In other words, this algorithm may assign two genes with low topological overlap to different modules even though their expression patterns are similar. Here, a novel gene module clustering algorithm for WGCNA is proposed. We develop a gene module clustering network (gmcNet), which simultaneously addresses single-level expression and topological overlap measure. The proposed gmcNet includes a "co-expression pattern recognizer" (CEPR) and "module classifier". The CEPR incorporates expression features of single genes into the topological features of co-expressed ones. Given this CEPR-embedded feature, the module classifier computes module assignment probabilities. We validated gmcNet performance using 4,976 genes from 20 native Korean cattle. We observed that the CEPR generates more robust features than single-level expression or topological overlap measure. Given the CEPR-embedded feature, gmcNet achieved the best performance in terms of modularity (0.261) and the differentially expressed signal (27.739) compared with other clustering methods tested. Furthermore, gmcNet detected some interesting biological functionalities for carcass weight, backfat thickness, intramuscular fat, and beef tenderness of Korean native cattle. Therefore, gmcNet is a useful framework for WGCNA module clustering.
BibTeX
@inproceedings{lee2022use,
  author = {Hyo-Jun Lee and Yoonji Chung and Ki Yong Chung and Young‐Kuk Kim and Jun Heon Lee and Yeong Jun Koh and Seung Hwan Lee},
  title = {{Use of a graph neural network to the weighted gene co-expression network analysis of Korean native cattle}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-13796-9},
}

Loading related content...

Figure