Deep Question Answering for protein annotation.

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Biomedical professionals face information overload from extensive literature.
• Standard search engines and question-answering (QA) systems struggle to efficiently retrieve specific information, especially for complex genomics questions.
• Existing QA systems often fail to extract answers requiring Gene Ontology (GO) concepts.

Purpose of the Study:

• To evaluate dictionary-based classifiers and a novel supervised classifier (GOCat) for extracting GO concepts from biomedical literature.
• To investigate the effectiveness of a deep QA approach that leverages curated biological data for inferring answers not explicitly stated.
• To address the limitations of current QA systems in handling complex genomics-related queries.

Main Methods:

• Comparison of two dictionary-based classifiers against a Gene Ontology (GO) supervised classifier (GOCat).
• Utilizing the GOA database to identify GO concepts annotated by curators for similar abstracts.
• Implementing a deep QA approach incorporating a classification step and curated data exploitation.
• Testing on a dataset of 100 retrieved abstracts per complex genomics question.

Main Results:

• Dictionary-based and redundancy-based QA approaches are relatively ineffective for complex genomics questions.
• The deep QA approach using GOCat significantly improves both the quantity and quality of extracted answers.
• A +100% improvement in both recall and precision was observed when using GOCat for complex answers like protein functional descriptions.

Conclusions:

• Standard QA methods are insufficient for complex genomics questions requiring Gene Ontology (GO) concepts.
• A deep QA approach, exemplified by GOCat, effectively utilizes curated biological data to infer answers.
• Supervised classification with curated data offers a substantial improvement in QA system performance for biomedical literature.