VAIV bio-discovery service using transformer model and retrieval augmented generation.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Computational Biology

Background:

• Advancements in AI, including Large Language Models (LLMs) and machine learning, are increasingly supporting biomedical knowledge discovery.
• Unstructured biomedical text, such as scientific literature, contains vast amounts of valuable information that is challenging to access and analyze.

Purpose of the Study:

• To introduce VAIV Bio-Discovery, a novel biomedical neural search service.
• To enhance knowledge discovery and document search capabilities for unstructured biomedical text.
• To facilitate the understanding of relationships between chemical compounds/drugs, genes/proteins, and diseases.

Main Methods:

• Developed a hybrid search engine combining neural search with BM25 probabilistic search.
• Utilized T5slim_dec, adapting the T5 (text-to-text transfer transformer) autoregressive generation for interaction extraction.
• Integrated Retrieval Augmented Generation (RAG) for summarizing search results based on natural language queries.

Main Results:

• The system effectively handles information on chemical compounds/drugs, genes/proteins, and diseases, including their interactions (e.g., drug-target, drug-drug, drug-disease).
• Four search options are provided: basic, entity and interaction search, and natural language search.
• The T5slim_dec model aids in interpreting research findings through result summarization.

Conclusions:

• The VAIV Bio-Discovery system demonstrates improved understanding of context, semantics, and term relationships within documents, leading to enhanced search accuracy.
• This novel service contributes to the biomedical field by providing improved access to and discovery of relevant knowledge.