MultiSourcDSim: an integrated approach for exploring disease similarity

Lei Deng, Lei Deng 0002, Danyi Ye, Junmin Zhao, Jingpu ZhangBmc Medical Informatics Decis. mak 2019 · DOI

TL;DR

This paper introduces MultiSourcDSim, a novel method that integrates multiple data sources (gene-disease, GO, and symptom associations) to more accurately predict disease similarity. By building a comprehensive disease similarity network, MultiSourcDSim outperforms existing methods and reveals latent relationships between diseases, aiding in understanding disease mechanisms and identifying potential therapeutic targets.

What Problem Does This Paper Address?

Understanding the relationships between diseases is crucial for advancing medical research and clinical practice. Identifying similar diseases can shed light on common pathogenic mechanisms, potentially leading to the repurposing of existing drugs or the development of new treatments. While numerous methods exist for calculating disease similarity, many rely on a single data source, limiting their accuracy and comprehensiveness. Diseases can be related through various biological pathways, and relying on just one type of data (e.g., gene associations) might miss crucial connections, especially for diseases with less well-understood genetic links. This paper addresses the challenge of improving disease similarity prediction by integrating diverse biological data sources to create a more robust and accurate representation of disease relationships.

Methods

The MultiSourcDSim approach involves three main steps:

  1. Constructing Disease Similarity Networks: Three distinct similarity networks are built, each based on a different data source:
  2. Gene-Disease Associations (GDN): Links diseases based on shared associated genes.
  3. GO Biological Process-Disease Associations (BPDN): Connects diseases through shared Gene Ontology (GO) biological processes.
  4. Symptom-Disease Associations (SDN): Relates diseases based on shared symptoms. The similarity within each network is calculated using Lin's method, which leverages information theory to measure semantic similarity based on common ancestors and individual term information content.

  5. Integrating Networks and Learning Representations: To combine the information from these three networks, the researchers employ Diffusion Component Analysis (DCA). This involves:

  6. Random Walk with Restart (RWR): Applied to each network to capture the topological structure and generate "diffusion states" for each disease node.
  7. Low-Dimensional Representation: A polynomial logistic model is used to approximate these high-dimensional diffusion states with lower-dimensional vectors. This step aims to capture the essential topological patterns while reducing noise and dimensionality. The objective function is minimized using KL-divergence.
  8. Calculating Disease Similarity: After obtaining the low-dimensional feature vectors for each disease from the integrated network, the similarity between any two diseases is calculated using the cosine similarity score between their respective vectors.

Datasets

The study utilizes several datasets:

  • CTD's MEDIC Disease Vocabulary: Used as the primary criterion for describing diseases, this vocabulary is a modified subset of the U.S. National Library of Medicine's MeSH (Medical Subject Headings) and includes genetic disorders from OMIM.
  • Gene-Disease Associations: A dataset containing 3,125,954 associations between 3,254 disease terms.
  • GO Biological Process-Disease Associations: A dataset with 668,760 associations involving 5,720 disease terms.
  • Symptom-Disease Associations: A dataset of 80,638 associations for 4,040 diseases, collected from a previous study. Disease names were mapped to MeSH identifiers.

After data integration and screening, a total of 8,126 diseases were extracted.

The paper also refers to a benchmark set of 40 highly similar disease pairs, derived from previous work and excluding cancers. This set is used for evaluating the performance of MultiSourcDSim against other methods.

Key Insights and Ideas

  • Synergy of Multiple Data Sources: The core insight is that integrating diverse biological data (genes, GO processes, symptoms) provides a more comprehensive and accurate picture of disease similarity than relying on any single source.
  • Capturing Latent Relationships: The integrated network constructed by MultiSourcDSim is capable of uncovering subtle, "latent" relationships between diseases that might be missed by single-source methods.
  • Robustness to Overfitting: The method demonstrates stability across a range of feature dimensions, suggesting it is robust and less prone to overfitting.
  • MeSH Category Correlation: The study confirms that disease pairs belonging to the same MeSH category exhibit significantly higher similarity scores, validating the biological relevance of the computed similarities.

Key Results

  • Superior Performance: MultiSourcDSim achieved the highest Area Under the ROC Curve (AUC) of 0.905 on the benchmark set, outperforming three other methods (HSDN: 0.818, MimMiner: 0.836, MedNetSim: 0.854). It also achieved the best scores for Accuracy (ACC), F1-score, Matthew's Correlation Coefficient (MCC), Precision (PRE), and Sensitivity (SEN).
  • Network Integration Advantage: When comparing integrated networks (GDN, BPDN, SDN) against individual networks, MultiSourcDSim (which integrates all three) showed a significantly higher AUC (0.906) compared to GDN (0.771), BPDN (0.774), and SDN (0.797).
  • MeSH Category Validation: The experiment showed that disease pairs within the same MeSH category had significantly higher average similarity scores than those from different categories, supporting the biological interpretability of the results.
  • Top-Ranking Performance: MultiSourcDSim consistently identified more disease pairs from the benchmark set within the top-ranking disease pairs compared to other methods.

Limitations and Future Work

  • Data Completeness Bias: The authors acknowledge that integrating multiple data sources can still be subject to bias due to incomplete data.
  • Modularity of Diseases: Future work could explore the modular nature of diseases further, as the "modularity of each disease module" might provide additional insights.
  • Therapeutic Applications: The disease networks generated by this approach could be useful for predicting novel therapeutic applications of known compounds and inferring novel disease genes.
  • Integration of More Data: While MultiSourcDSim integrates three key data types, future research could explore incorporating even more diverse biological data sources.

Summary updated

BibTeX
@article{deng2019multisourcdsim,
  author = {Lei Deng and Lei Deng and Danyi Ye and Junmin Zhao and Jingpu Zhang},
  title = {{MultiSourcDSim: an integrated approach for exploring disease similarity}},
  journal = {Bmc Medical Informatics Decis. mak},
  year = {2019},
  doi = {10.1186/s12911-019-0968-8},
}
Figure