Exploring Biomolecular Literature with EVEX: Connecting Genes through Events, Homology, and Indirect Associations

Sofie Van Landeghem, Kai Hakala, Samuel Rönnqvist, Tapio Salakoski, Yves Van de Peer, Filip GinterAdv. Bioinformatics 2012 · DOI

TL;DR

This paper introduces EVEX, a web application that makes a massive dataset of over 21 million biomolecular events extracted from PubMed abstracts easily searchable. EVEX connects genes through direct and indirect relationships, leveraging homology and event generalization to help researchers discover novel biological hypotheses.

What Problem Does This Paper Address?

The life sciences are drowning in published literature, making it incredibly difficult for researchers to keep up with existing knowledge, even within their narrow fields. While text mining offers a solution, extracting detailed, actionable information from this vast corpus has been a significant challenge. Specifically, previous approaches often extracted only basic, untyped relationships between biomolecular entities (like genes and proteins). This paper tackles the need for a more comprehensive understanding by extracting detailed biomolecular events (like phosphorylation or regulation) and making this information accessible through an intuitive interface.

Methods

The core of EVEX is built upon the Turku Event Extraction System (TEES), a state-of-the-art system that extracts detailed biomolecular events from text. The authors then enrich these "text-bound" events in several key ways:

  • Canonicalization: Gene and protein symbols are standardized to handle variations in naming.
  • Gene Family Generalization: Canonicalized symbols are further mapped to gene families using resources like HomoloGene and Ensembl. This allows for abstracting information across homologous genes and species.
  • Event Refinement: Chains of single-argument regulatory events are simplified to reduce complexity and improve aggregation.
  • Pairwise Abstraction: Events are analyzed to identify direct and indirect relationships between pairs of genes.
  • Indirect Association Discovery: EVEX identifies coregulators and common binding partners by combining information from multiple events, enabling hypothesis generation.
  • Confidence Scoring: A machine learning-based scoring system ranks extracted events by reliability, allowing users to prioritize high-confidence findings.

Datasets

The primary dataset used is the EVEX dataset, which contains over 21 million detailed biomolecular events extracted from approximately 6 million PubMed abstracts (covering titles and abstracts from 2009-2011). This dataset was built upon previous work that processed the 2009 PubMed distribution.

Key Insights and Ideas

  • Bridging the Gap: EVEX successfully translates complex text-mining outputs into a user-friendly web application, making detailed biomolecular event data accessible to a broader audience beyond NLP specialists.
  • Beyond Direct Links: The ability to discover indirect associations (like coregulation) is a powerful tool for hypothesis generation, revealing connections that might not be immediately obvious from individual studies.
  • Homology as a Bridge: Leveraging gene family information allows for cross-species and cross-organism inference, significantly expanding the utility of the extracted data.
  • Confidence is Key: Providing confidence scores for extracted events is crucial for users to assess the reliability of the information and prioritize their investigations.
  • Text Mining for Pruning: EVEX can help identify and filter out false positives from text mining results by comparing evidence and confidence scores.

Key Results

  • The EVEX dataset contains over 21 million detailed biomolecular events extracted from nearly 6 million PubMed abstracts.
  • The web application provides an intuitive interface for exploring these events, including direct and indirect associations between genes.
  • Confidence values for events range from "very low" to "very high," with high-confidence events achieving precision rates as high as 91.9% (though with lower recall).
  • Gene family generalization, particularly with the inclusion of Ensembl Genomes, significantly increases coverage, accounting for over 60% of gene symbol occurrences.
  • Event refinement reduces the number of distinct event structures by over 60%, simplifying analysis.

Limitations and Future Work

  • Manual Analysis Still Needed: While EVEX provides powerful tools, users still need to manually analyze some results, especially for "general sentences" that don't contain extracted events.
  • Semantic Coverage: A significant reason for interactions not being extracted is a lack of semantic coverage in the event extraction task itself, rather than a failure of the pipeline.
  • Future Work:
    • Expand event extraction to include full-text articles from PubMed Central.
  • Incorporate new information types like epigenetics data and non-causal entity relations.
  • Integrate gene normalization data for queries using specific gene/protein identifiers.
  • Develop a web service for programmatic access to the EVEX data, enabling bulk queries and result export for integration into other bioinformatics pipelines.

Summary updated

BibTeX
@article{landeghem2012exploring,
  author = {Sofie Van Landeghem and Kai Hakala and Samuel Rönnqvist and Tapio Salakoski and Yves Van de Peer and Filip Ginter},
  title = {{Exploring Biomolecular Literature with EVEX: Connecting Genes through Events, Homology, and Indirect Associations}},
  journal = {Adv. Bioinformatics},
  year = {2012},
  doi = {10.1155/2012/582765},
}
Figure