Gene prioritization and clustering by multi-view text mining

Shi Yu, Léon-Charles Tranchevent, Bart De Moor, Yves MoreauBMC Bioinformatics 2010 · DOI

TL;DR

This paper introduces a "multi-view text mining" approach to improve the identification of candidate genes associated with diseases. By integrating information from multiple controlled vocabularies (views) and applying dimensionality reduction techniques, they achieve significantly better performance in both gene prioritization and clustering tasks compared to using single vocabularies.

What Problem Does This Paper Address?

Biologists often use text mining to sift through vast amounts of biomedical literature (like MEDLINE) to identify genes potentially linked to specific diseases. This is crucial for experimental validation. However, different text mining models, often based on different controlled vocabularies (CVs), yield varying results. The challenge lies in effectively combining these diverse "views" of the data to achieve more accurate and reliable gene identification, ensuring the integrated model outperforms any single model. This paper tackles the problem of how to best fuse these multiple perspectives to enhance gene prioritization and clustering.

Methods

The core of the proposed method is multi-view text mining. This involves:

  • Controlled Vocabulary Selection: Nine well-known bio-ontologies (e.g., Gene Ontology, MeSH, OMIM) were used to create distinct "views" of the biomedical literature. Each CV provides a different semantic perspective.
  • Text Mining and Profile Creation: MEDLINE abstracts and titles related to human genes were indexed using these CVs. This process generated "gene-by-term" profiles, where each gene is represented by its association with terms from a specific CV.
  • Dimensionality Reduction: To handle the high dimensionality of gene-by-term profiles, two techniques were employed:
  • Latent Semantic Indexing (LSI): A standard information retrieval technique to reduce the number of features while preserving essential information.
  • Subset CVs: Selecting terms from specific sub-ontologies within a larger CV to create more focused views.
  • Data Fusion (Integration): Multiple gene-by-term profiles (views) were combined using various machine learning algorithms. The paper explored both ensemble methods (combining rankings) and kernel fusion methods (combining similarity matrices). A key aspect is the "integration after splitting" strategy, where individual CVs are first processed and then integrated, rather than merging all terms into one large vocabulary.
  • Tasks: The integrated approach was evaluated on two fundamental tasks:
  • Gene Prioritization: Ranking candidate genes based on their likelihood of being involved in a disease.
  • Gene Clustering: Grouping genes based on their associations with diseases.

Datasets

  • MEDLINE Repository: Used as the primary source of biomedical literature. The study utilized a subset of publications related to human genes (as of December 2008).
  • Human Disease Benchmark Data Set (Endeavour): A dataset consisting of 620 relevant genes categorized into 29 genetic diseases. This dataset was used for both prioritization and clustering evaluations.

Key Insights and Ideas

  • The Power of Multiple Perspectives: Combining information from different controlled vocabularies (views) significantly improves the accuracy of gene identification tasks. No single vocabulary is universally optimal, making a multi-view approach robust.
  • "Integration After Splitting" is Key: Merging all terms from multiple vocabularies into a single large vocabulary ("merge-9") performed worse than integrating individual CVs. This suggests that preserving the distinct semantic contributions of each CV is crucial.
  • Dimensionality Reduction Enhances Fusion: Techniques like LSI and using subset CVs not only reduce computational burden but also improve the performance of data fusion methods.
  • Robustness to Irrelevant Views: The multi-view approach is less sensitive to the inclusion of less relevant or noisy views, leading to more stable and reliable results. This is a significant advantage for practical exploratory analysis.

Key Results

  • Prioritization: The best performance was achieved by combining 9 LSI-reduced CV profiles using a 1-SVM kernel fusion method, resulting in an Error of AUC of 0.0335. This was roughly half the error of the best single CV profile (LDDB, 0.0792).
  • Clustering: The best clustering performance was obtained by combining 9 complete CV profiles using Ward linkage (RI = 0.8236, NMI = 0.6015). Other multi-view integration methods also outperformed the best single CV (LDDB, RI = 0.7586, NMI = 0.5290).
  • Comparison to Existing Systems: The proposed multi-view approach showed better performance than existing systems like Endeavour on the same benchmark dataset.
  • Effectiveness of View Integration: Integrating multiple views consistently improved performance across various integration strategies and numbers of views, demonstrating the robustness of the multi-view approach.

Limitations and Future Work

  • Interpretation of LSI: The authors note that LSI's latent factors are not easily interpretable in terms of specific terms affecting prioritization.
  • Cluster Number Selection: For K-means based clustering, selecting the optimal number of clusters can be challenging, especially with heterogeneous data sources.
  • Overlapping Genes: The current approach primarily handles distinct gene-disease associations. Future work could explore "soft clustering" techniques to account for genes relevant to multiple diseases or functions.
  • Automated Gene Name Recognition: The current method relies on curated GeneRIF for gene-publication mapping. Future work could incorporate NLP to automatically detect gene names and build more dynamic gene profiles.
  • Exploring More Parameters: The authors suggest that multi-view text mining could be extended to incorporate other parameters or data types, such as gene interaction networks.

Summary updated

BibTeX
@article{yu2010gene,
  author = {Shi Yu and Léon-Charles Tranchevent and Bart De Moor and Yves Moreau},
  title = {{Gene prioritization and clustering by multi-view text mining}},
  journal = {BMC Bioinformatics},
  year = {2010},
  doi = {10.1186/1471-2105-11-28},
}
Figure