Large Language Models for Accessible Reporting of Bioinformatics Analyses in Interdisciplinary Contexts.

Area of Science:

• Bioinformatics
• Computational Biology
• Scientific Communication

Background:

• Interdisciplinary collaboration between health/life scientists and quantitative scientists is vital for data analysis.
• Miscommunication frequently hinders the interpretation of complex scientific results.
• Large Language Models (LLMs) show potential for bridging communication gaps in scientific research.

Purpose of the Study:

• To benchmark the interpretative capabilities of state-of-the-art LLMs in real-world bioinformatics analyses.
• To evaluate LLM performance using both automated and human assessment frameworks.
• To determine the extent to which LLMs can facilitate cross-disciplinary understanding in science.

Main Methods:

• Benchmarking four leading LLMs: GPT-4o, o1, Claude 3.7 Sonnet, and Gemini 2.0 Flash.
• Utilizing automated assessments with multiple-choice questions based on Bloom's taxonomy.
• Employing human evaluations where scientists scored LLM summaries on factual consistency, safety, comprehensiveness, and coherence.

Main Results:

• LLMs produced generally readable and safe summaries, useful for initial translation of technical analyses.
• LLMs frequently misinterpreted data visualizations and generated overly verbose outputs.
• LLMs rarely offered novel scientific insights beyond the provided analytical data.

Conclusions:

• LLMs are valuable tools for facilitating interdisciplinary communication in science.
• LLMs should be used to aid, not replace, domain expertise and human oversight.
• Human expertise remains essential for ensuring accuracy, interpretative depth, and generating novel scientific discoveries.