KDGene: knowledge graph completion for disease gene prediction using interactional tensor decomposition

Xinyan Wang, Kuo Yang, Ting Jia, Fanghui Gu, Chongyu Wang, Kuan Xu, Zixin Shu, Jianan Xia, Qiang Zhu, Xuezhong ZhouBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces KDGene, a novel framework for predicting disease-associated genes by leveraging a comprehensive biological knowledge graph and an interactional tensor decomposition method. KDGene significantly improves upon existing methods by better capturing complex relationships within biological data, leading to more accurate identification of candidate genes.

What Problem Does This Paper Address?

Accurately identifying genes linked to specific diseases is crucial for understanding disease mechanisms and developing precision medicine. Traditional experimental methods are time-consuming and labor-intensive. While computational approaches, particularly those using machine learning and deep learning on biological networks, have advanced, they often overlook the intricate relationships within biological knowledge graphs. Existing knowledge graph embedding (KGE) methods, while effective in general domains, often fall short when applied to the complex, domain-specific data of life sciences. This gap limits their ability to accurately predict disease genes.

Methods

The core of KDGene is a knowledge graph completion (KGC) framework built upon tensor decomposition. The authors first construct a large-scale biological knowledge graph by integrating various relation types (e.g., Disease-Symptom, Protein-Protein Interactions, Drug-Disease) from multiple biological databases.

The novelty lies in the interaction module integrated into a tensor decomposition model (specifically, CP-N3). This module uses a gating mechanism, inspired by Long Short-Term Memory (LSTM) networks, to allow for more nuanced interaction between entity and relation embeddings. This allows the model to better perceive and interpret the complex relationships within biological knowledge, leading to more precise representations of entities and relations. Unlike traditional KGE methods that require fixed embedding dimensions, KDGene's interaction module allows for flexible, unequal dimensions for entities and relations, enhancing operability and flexibility.

The KGC task is framed as predicting missing links in the knowledge graph, specifically identifying genes associated with a given disease. The model is trained using a standard data augmentation technique and a 1-N scoring approach, with a loss function incorporating N3 regularization.

Datasets

The primary dataset used is the DisGeNet database, which contains curated disease-gene associations. This dataset is used for training and evaluating the model using 10-fold cross-validation. The authors also construct a comprehensive biological knowledge graph by integrating data from various sources, including:

  • SymMap (Disease-Symptom relations)
  • SIDER (Drug-Disease relations)
  • STRING (Protein-Protein Interactions)
  • STITCH (Drug-Protein relations)
  • Gene Ontology (GO) (Protein-GO relations)
  • KEGG (Protein-Pathway relations)

Key Insights and Ideas

  • The power of integrated biological knowledge: Simply using disease-gene facts is insufficient. Incorporating diverse biological relations (like symptoms, protein interactions, and drug associations) significantly enhances prediction accuracy.
  • Interaction module is key: The proposed interaction module, inspired by LSTM gates, is crucial for improving how KDGene understands the interplay between entities and relations. This is a significant step beyond standard tensor decomposition.
  • Flexibility in embedding dimensions: Allowing different embedding dimensions for entities and relations makes the model more adaptable and potentially more efficient for biological data.
  • Not all relations are equal: The study highlights that some relations (like Disease-Symptom) are more informative for disease gene prediction than others (like Protein-Protein Interactions alone), suggesting a need for careful KG construction.
  • Reliability of high-confidence data: Using higher confidence scores for protein-protein interactions leads to better performance, emphasizing the importance of data quality.

Key Results

  • Superior performance: KDGene consistently outperforms both state-of-the-art Disease Gene Prediction (DGP) methods and existing Knowledge Graph Embedding (KGE) baselines.
  • Significant improvements: Compared to DGP baselines, KDGene shows an average improvement of 16.59%. When compared to KGE baselines, the average improvement is 21.96%.
  • Impact of the interaction module: KDGene achieves an impressive average improvement of 43.77% over its baseline CP-N3 (which lacks the interaction module), demonstrating the module's critical contribution.
  • Robustness on unseen data: KDGene shows promising performance in predicting associations for diseases not explicitly present in the training set, indicating good generalization capabilities.
  • Case study validation: Functional homogeneity analysis and case studies on diabetes mellitus and atrophic gastritis show that KDGene can identify novel, biologically relevant candidate genes with strong literature and database support.

Limitations and Future Work

  • Computational cost: While not explicitly detailed as a limitation, the paper mentions that computational cost and complexity were analyzed in supplementary materials, suggesting it's a consideration.
  • Data availability and complexity: The authors acknowledge the rapid development of biological databases and the emergence of new works, which pose challenges for direct comparison. They also note that incorporating too many similar entities (like in PPIs) can sometimes introduce noise and reduce effectiveness.
  • Future work: The authors aim to incorporate insights from advanced methodologies and continually optimize their model to improve predictive capabilities, especially in the broader context of computational biology. They also suggest that the framework is scalable and can accommodate new relation types as biological databases grow.

Summary updated

BibTeX
@article{wang2024kdgene,
  author = {Xinyan Wang and Kuo Yang and Ting Jia and Fanghui Gu and Chongyu Wang and Kuan Xu and Zixin Shu and Jianan Xia and Qiang Zhu and Xuezhong Zhou},
  title = {{KDGene: knowledge graph completion for disease gene prediction using interactional tensor decomposition}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae161},
}
Figure