Multi-view Contrastive Learning Hypergraph Neural Network for Drug-Microbe-Disease Association Prediction

Luotao Liu, Feng Huang, Xuan Liu, Zhankun Xiong, Menglu Li, Congzhi Song, Wen ZhangIJCAI 2023 · DOI

Introduction

Understanding the complex interplay between drugs, microbes, and diseases (DMD) is a cornerstone of modern precision medicine. While traditional computational methods focus on pairwise associations (e.g., drug-microbe or microbe-disease), real-world biological systems involve higher-order interactions. Predicting these triple-wise associations is challenging due to the extreme sparsity of confirmed data; in current datasets, only 0.11% of possible DMD combinations are verified. This "data desert" often leads to overfitting in standard deep learning models.

To address these challenges, researchers have introduced the Multi-view Contrastive Learning Hypergraph Neural Network (MCHNN). This approach leverages hypergraph structures to capture high-order dependencies and employs a novel contrastive learning strategy to regularize the model against sparse data.

Methodology

The MCHNN framework operates through a four-stage pipeline designed to extract robust representations from heterogeneous biological data.

The MCHNN architecture: (1) Feature encoding via GCN/FCN, (2) Hypergraph representation learning, (3) Multi-view contrastive learning for data augmentation, and (4) MLP-based association scoring.
  1. Feature Encoding: Molecular structures of drugs (SMILES) are processed using a Graph Isomorphism Network (GIN), while microbes and diseases are encoded via Fully Connected Networks (FCN) based on taxonomic and MeSH similarity matrices.
  2. Hypergraph Representation Learning: The model constructs a hypergraph where nodes represent entities and hyperedges represent known DMD triplets. A Hypergraph Convolutional Network (HGCN) propagates information across these complex connections.
  3. Multi-view Contrastive Learning: To combat sparsity, the model generates four "augmented" views by perturbing hyperedges (e.g., swapping a single entity type). By training the model to distinguish real biological triplets from these perturbed versions, it learns more discriminative features.
  4. Association Prediction: Finally, a Multi-Layer Perceptron (MLP) aggregates the learned embeddings to calculate a DMD association score.

Experimental Results

MCHNN was evaluated against several state-of-the-art baselines, including tensor factorization (CP, Tucker) and existing hypergraph models like HypergraphSynergy.

Quantitative Performance

The model demonstrated superior predictive accuracy across multiple scenarios, achieving an average Hit@5 of 0.9340 and an NDCG@5 of 0.9280. On independent test sets, MCHNN consistently outperformed the best-performing baselines by margins ranging from 0.91% to 6.20% in average Hit@5 metrics.

Embedding Quality

Qualitative analysis using t-SNE visualizations confirms that MCHNN produces superior node embeddings. Compared to other deep learning models, MCHNN creates tighter and more distinct clusters for positive and negative associations.

As shown in the visualization, MCHNN achieves a silhouette score of approximately 0.5287, significantly higher than HypergraphSynergy (0.3864), indicating that its contrastive learning strategy effectively captures the underlying biological patterns despite data sparsity.

Conclusion

MCHNN represents a significant advancement in predicting complex biological associations. By moving beyond pairwise modeling and utilizing contrastive learning as a powerful regularizer, the model provides a robust tool for identifying potential therapeutic combinations. Future work may involve incorporating more diverse data types, such as genomic sequences, and predicting the specific directionality or type of microbial abundance changes induced by drugs in specific disease contexts.

Summary updated

BibTeX
@inproceedings{liu2023multi,
  author = {Luotao Liu and Feng Huang and Xuan Liu and Zhankun Xiong and Menglu Li and Congzhi Song and Wen Zhang},
  title = {{Multi-view Contrastive Learning Hypergraph Neural Network for Drug-Microbe-Disease Association Prediction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/537},
}
Figure