Building a search tool for compositely annotated entities using Transformer-based approach: Case study in...

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Transformer-based models like BERT and GPT excel in Natural Language Processing (NLP) tasks.
• Their effectiveness stems from large pre-trained datasets and contextual word understanding.
• Applying these NLP techniques to information retrieval can boost efficiency and performance.

Purpose of the Study:

• To implement a BERT-based method (CASBERT) for creating a search tool.
• To enable searching through compositely-annotated biosimulation models from the Physiome Model Repository (PMR).
• To facilitate the retrieval of specific entities within biosimulation models.

Main Methods:

• Developed embeddings for compositely-annotated entities (constants, variables) in CellML models.
• Vertically and efficiently combined low-level entity embeddings to create higher-level entity embeddings.
• Implemented a BERT-based approach for searching various model levels (components, models, images, simulation setups).

Main Results:

• Successfully created a BERT-based search tool named Biosimulation Model Search Engine (BMSE).
• Demonstrated the ability to search for both low-level (constants, variables) and high-level (components, models) entities.
• The method allows for efficient retrieval of specific information within complex biosimulation models.

Conclusions:

• The CASBERT implementation provides an effective and efficient search tool for biosimulation models.
• The generalizable approach can be adapted for other ontology-annotated data, such as SBML models.
• This work advances information retrieval in computational biology by leveraging advanced NLP techniques.