Detecting experimental techniques and selecting relevant documents for protein-protein interactions from biomedical...

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• BioCreative III contest focused on selecting relevant articles for protein-protein interaction (PPI) curation and identifying experimental techniques.
• Two sub-tasks were defined: Article Classification Task (ACT) for document selection and Interaction Method Task (IMT) for technique recognition.

Purpose of the Study:

• To develop and compare classification-based methods for ACT and IMT using contextual and external knowledge features.
• To evaluate the performance of novel and conventional approaches for PPI article curation and method identification.

Main Methods:

• Proposed classification methods utilizing rich contextual features and external knowledge sources.
• Developed a novel approach for IMT classifying pair-wise relations between text phrases and interaction methods.
• Explored combining novel and multi-label document classification methods for IMT.
• For ACT, employed classifiers using automatically detected named entities and linguistic information.

Main Results:

• The novel IMT method achieved an F1 score of 64.49% on the development dataset.
• One IMT method achieved the best performance among all participants in BioCreative III based on F1 score, MCC, and AUC iP/R.
• The best ACT classifier ranked second by AUC iP/R and was competitive on other metrics.

Conclusions:

• A novel binary classification approach showed promise for IMT, with best performance achieved by combining it with a multi-class, multi-label classifier.
• ACT system performance was encouraging, with contextual words around named entities and MeSH headings identified as key features.