BioNorm: deep learning-based event normalization for the curation of reaction databases

Area of Science:

• Biomedical text mining
• Bioinformatics
• Computational biology

Background:

• Biochemical reactions (bio-events) describe entity relationships.
• Identifying bio-events from literature is a focus of text mining.
• Normalizing extracted bio-events to curated databases is underexplored, hindering network construction.

Purpose of the Study:

• To propose BioNorm, a novel method for normalizing bio-events from scientific literature to curated reaction databases like IntAct.
• To address the challenge of disambiguating bio-events and enabling the creation of comprehensive biological networks.

Main Methods:

• BioNorm treats bio-event normalization as a paraphrase identification problem.
• It converts database entries into natural language statements.
• A long short-term memory recurrent neural network (LSTM) predicts semantic similarity between literature statements and database entries.

Main Results:

• The BioNorm method achieves an F-score of 0.87 in normalizing event-containing statements.
• Experiments utilized molecular interaction data from the IntAct database.
• This represents the first attempt at bio-event normalization in biomedical text mining.

Conclusions:

• BioNorm effectively normalizes bio-events from scientific literature to curated databases.
• The method enhances the potential for building interconnected, biologically meaningful networks.
• The approach offers a significant advancement in biomedical text mining capabilities.