BioConceptVec

BioConceptVec generates concept embeddings for primary biological concepts mentioned in biomedical literature to capture semantic relatedness among genes, mutations, proteins, drugs, and other entities for downstream tasks such as protein–protein interaction prediction and literature-based discovery.

Key Features:

• Extensive Coverage: Contains over 400,000 biomedical concepts extracted from approximately 30 million PubMed abstracts.
• Advanced Named-Entity Recognition (NER): Utilizes high-performance machine learning-based NER tools for identification and normalization of biological concepts within the literature.
• Diverse Machine Learning Models: Trains concept embeddings using four distinct machine learning models to capture complex semantic relationships.
• Comprehensive Evaluation: Underwent intrinsic evaluation on 17 million instances and extrinsic evaluation on 8 million instances across three tasks.

Scientific Applications:

• Protein-Protein Interaction Prediction: Embeddings are used to improve prediction of protein–protein interactions.
• Drug-Gene and Gene-Gene Interaction Identification: Facilitates identification of drug–gene and gene–gene interactions, with intrinsic evaluation demonstrating superior performance relative to existing concept embeddings.
• Drug-Drug Interaction Extraction: Supports extraction of drug–drug interaction data for pharmacological research.
• Literature-Based Discovery: Captures semantic relatedness among biomedical entities to support literature-based discovery and hypothesis generation.

Methodology:

Concept recognition from approximately 30 million PubMed abstracts identified and normalized over 400,000 biomedical concepts using advanced machine learning-based NER tools. The identified concepts were trained to produce vector representations using four distinct machine learning models. The embeddings underwent intrinsic evaluation on 17 million instances and extrinsic evaluation on 8 million instances across three tasks.