WSGMB: weight signed graph neural network for microbial biomarker identification

Shuheng Pan, Xinyi Jiang, Kai ZhangBriefings in Bioinformatics 2024 · DOI

Abstract

WSGMB is a novel computational framework designed to identify disease-related microbial biomarkers by analyzing changes in the gut microbiome's interaction network. By representing microbial interactions as weighted, signed graphs and employing a specialized Graph Neural Network (GNN), WSGMB identifies key bacteria whose ecological roles shift significantly between healthy and diseased states. This approach addresses the limitations of traditional abundance-based methods, offering superior performance in both disease classification and biomarker discovery.

Introduction

The human gut microbiome is a complex ecosystem where microbial taxa interact through symbiosis and competition. Traditional biomarker identification typically relies on differential abundance—comparing the raw counts of bacteria between healthy and sick individuals. However, abundance data is often noisy and fails to capture the systemic shifts in microbial communities. WSGMB shifts the focus to the co-occurrence network, identifying "hub" microbes whose connectivity patterns drive the transition to disease.

Methodology

The WSGMB framework operates through three primary stages: network construction, graph classification via a custom GNN, and a node "kick-out" strategy for biomarker identification.

Overview of the WSGMB framework, illustrating the pipeline from OTU tables to biomarker identification.

Network Construction

Using the SparCC algorithm, the framework builds weighted signed graphs from microbial abundance data. Nodes represent taxa, while edges represent correlations. The "sign" of an edge indicates potential symbiosis (positive) or competition (negative), and the "weight" reflects the interaction strength. To ensure robustness across varying sample sizes, the model utilizes dynamic thresholding based on average dataset correlations.

WSGCNN Architecture

The core of the framework is the Weight Signed Graph Convolutional Neural Network (WSGCNN). Unlike standard GNNs, WSGCNN processes positive and negative neighbors using distinct aggregators to preserve the semantic meaning of different interaction types. It incorporates Graph Self-Adaptive Pooling (GSAPool) to rank node importance based on both topology and features.

Detailed architecture of the WSGCNN model, showing message passing, feature concatenation, and pooling layers.

Node Knock-out Strategy

To identify biomarkers, the model performs sensitivity analysis. By systematically removing a microbe and its associated edges (the "kick-out"), the system measures the resulting drop in classification accuracy. Microbes that cause the most significant change in prediction between healthy and diseased states are flagged as high-importance biomarkers.

Results and Validation

The model was evaluated using data from the gutMDisorder database, specifically focusing on Colorectal Cancer (CRC) and Crohn’s Disease (CD).

  • Classification Accuracy: WSGCNN achieved near-perfect performance, reaching up to 100% AUC in distinguishing between healthy and diseased microbial networks.
  • Biomarker Identification: The framework outperformed state-of-the-art methods like CACONET and NetMoss, achieving AUCs over 70% for CRC and 65% for CD.
  • Biological Relevance: Validation against existing literature confirmed the identification of known beneficial genera, such as Lachnospira, which is recognized for producing short-chain fatty acids that help prevent colorectal cancer.

Conclusion and Future Work

WSGMB demonstrates that microbial interactions are more stable and informative indicators of disease than raw abundance. While current limitations include a reliance on simple node features (e.g., node degree) and single-node sensitivity analysis, future work aims to incorporate multi-omic data and explore multi-node interaction groups to further refine biomarker discovery in complex disease states.

Summary updated

BibTeX
@article{pan2024wsgmb,
  author = {Shuheng Pan and Xinyi Jiang and Kai Zhang},
  title = {{WSGMB: weight signed graph neural network for microbial biomarker identification}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbad448},
}
Figure