Steering veridical large language model analyses by correcting and enriching generated database queries: first steps...

Area of Science:

• Genomics and Bioinformatics
• Artificial Intelligence in Scientific Research

Background:

• Large language models (LLMs) possess factual knowledge but often lack completeness and retrieval accuracy, especially in rapidly evolving scientific fields like genomics.
• Existing LLMs, such as ChatGPT, exhibit limitations as bioinformatics assistants, including poor data retrieval, hallucination, and incorrect sequence manipulation.

Purpose of the Study:

• To address the limitations of LLMs in scientific domains by developing a system that grounds LLM outputs in current, authoritative data.
• To enhance LLM capabilities for data analysis in genomics and bioinformatics through improved factuality and instruction following.

Main Methods:

• Introduction of NagGPT, a middleware tool designed to interface between LLMs and databases, managing queries and responses.
• Development of a companion OpenAI custom GPT, Genomics Fetcher-Analyzer, which directs ChatGPT to generate and execute Python code for bioinformatics tasks using data from multiple genomics databases.
• Implementation of strategies to mitigate challenges such as code generation issues, identifier confusion, and data hallucination.

Main Results:

• Demonstration of NagGPT's effectiveness in bridging LLM knowledge gaps and facilitating database API usage.
• Successful execution of bioinformatics tasks by ChatGPT, guided by Genomics Fetcher-Analyzer and powered by dynamically retrieved data.
• Partial mitigation of issues related to code-data interaction, identifier ambiguity, and LLM hallucination.

Conclusions:

• The proposed system, incorporating NagGPT and Genomics Fetcher-Analyzer, offers a viable approach to augment LLMs as specialized bioinformatics assistants.
• Findings suggest pathways to enhance the factual accuracy and instruction-following capabilities of unmodified LLMs in scientific contexts.