BioConceptVec: Creating and evaluating literature-based biomedical concept embeddings on a large scale.

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• Biomedical literature contains vast information on genes and mutations.
• Capturing semantic relatedness of biological entities is crucial for applications like protein-protein interaction prediction.
• Existing biomedical concept embeddings often use suboptimal tools and have limited evaluation and availability.

Purpose of the Study:

• To develop a novel, high-performance biomedical concept embedding, BioConceptVec.
• To address limitations of existing concept embeddings regarding NER tools, evaluation scale, and availability.
• To provide a comprehensive and publicly accessible resource for biomedical concept representation.

Main Methods:

• Utilized high-performance machine learning-based named-entity recognition (NER) tools for concept identification.
• Trained BioConceptVec embeddings using four different machine learning models on approximately 30 million PubMed abstracts.
• Conducted extensive intrinsic and extrinsic evaluations on over 25 million instances across nine independent datasets.

Main Results:

• BioConceptVec covers over 400,000 biomedical concepts, making it one of the largest publicly available embeddings.
• Achieved significantly better performance than existing embeddings in intrinsic evaluations for identifying related concepts.
• Demonstrated substantial improvements in downstream bioinformatics studies and biomedical text-mining applications through extrinsic evaluations.

Conclusions:

• BioConceptVec represents a significant advancement in biomedical concept embeddings.
• The comprehensive evaluation validates its superior performance and utility.
• Public availability of BioConceptVec and benchmarking datasets facilitates further research in bioinformatics and text mining.