ONBRIRES: Ontology-Based Biological Relation Extraction System

Minlie Huang, Xiaoyan Zhu, Shilin Ding, Hao Yu, Ming Li

Abstract

The rapid growth of biomedical literature necessitates automated systems for knowledge extraction. ONBRIRES (ONtology-Based BIological Relation Extraction System) is a unified framework designed to extract diverse biological relations, such as protein-protein interactions and gene-disease correlations, from unstructured MEDLINE abstracts. By integrating external ontologies with a novel two-level hierarchical pattern learning algorithm, the system achieves enhanced portability and a superior balance between precision and recall compared to traditional, single-purpose extraction tools.

Introduction

The explosion of biological research has resulted in a vast amount of critical knowledge being "locked" within millions of text-based scientific papers. While structured databases like DIP or KEGG exist, manual curation is increasingly impractical. Existing Information Extraction (IE) tools often suffer from a lack of portability, being tailored to specific relation types, and frequently require extensive manual labor for rule creation. ONBRIRES addresses these gaps by providing an automated, ontology-aware system capable of learning extraction patterns across different biological domains.

System Architecture

The ONBRIRES architecture is built upon three main pillars: ontology integration, hierarchical pattern learning, and weighted pattern matching.

The ONBRIRES architecture overview, showing the flow from entity pair input to relation extraction.

The system integrates four major external ontologies—Gene Ontology (GO), MeSH, LocusLink, and OMIM—into a uniform format. This integration allows for entity expansion into "synsets" (sets of synonyms), enabling the system to recognize biological entities regardless of the specific nomenclature used by authors.

Hierarchical Pattern Learning

A core innovation of ONBRIRES is its two-level pattern learning hierarchy, which moves beyond simple keyword matching.

The pattern hierarchy illustrating the relationship between syntactic, token, and instance levels.

The hierarchy consists of:

  • Token Patterns: These focus on specific keywords (e.g., "interacts with"). While highly precise, they often suffer from low coverage.
  • Syntactic Patterns: These utilize Part-of-Speech (POS) tags (e.g., Noun-Phrase + Verb + Preposition + Noun-Phrase) to capture broader linguistic structures, helping identify relations even when specific vocabulary is novel.

By organizing patterns from specific instances to general syntactic rules, the system can fall back on broader patterns when specific token matches fail, significantly boosting recall.

Experimental Setup

The system was evaluated on two distinct biological tasks:

  1. Protein-Protein Interaction (PPI): Trained on the MIPS database and evaluated using the GENIA corpus.
  2. Gene-Disease Correlation: Utilized data from the OMIM (Online Mendelian Inheritance in Man) database for both training and evaluation.

Results and Discussion

The evaluation compared "Token-only" matching against a combined "Token + Syntactic" approach. On the GENIA (PPI) dataset, the combined approach reached an F-score of 0.306. Performance was notably higher on the OMIM (Gene-Disease) dataset, achieving an F-score of 0.58.

A key finding was the "convergence" effect: at very high matching thresholds, both methods perform similarly as only perfect token matches are accepted. However, at lower thresholds, syntactic patterns provide a significant performance boost by capturing valid relations expressed through varied vocabulary. Furthermore, the use of ontologies makes the system "semantic-aware," treating different strings representing the same concept (e.g., "p53" and "tumor protein p53") as equivalent.

Conclusion and Future Work

While ONBRIRES demonstrates improved portability and performance, several challenges remain. Future work will focus on improving pattern quality to reduce false positives, handling complex grammatical structures like appositives, and resolving entity ambiguity, particularly between genes and proteins that share identical names.

Summary updated

BibTeX
@inproceedings{huang2006onbrires,
  author = {Minlie Huang and Xiaoyan Zhu and Shilin Ding and Hao Yu and Ming Li},
  title = {{ONBRIRES: Ontology-Based Biological Relation Extraction System}},
  booktitle = {APBC},
  year = {2006},
  url = {https://openreview.net/forum?id=eSla276eUe},
}
Figure