Reflections on the use of LLMs for cell annotation

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Single-cell and spatial transcriptomics generate large datasets requiring robust cell type annotation.
• Large Language Models (LLMs) are emerging as powerful tools for biological data interpretation.
• The AICellType platform offers LLM-based annotation but relies on proprietary models.

Purpose of the Study:

• To evaluate the AICellType platform for LLM-based cell type annotation.
• To raise concerns regarding the reliance on proprietary commercial LLMs in biomedical research.
• To advocate for the adoption of open-source alternatives and enhanced system features.

Main Methods:

• Systematic benchmarking of the AICellType platform.
• Critical analysis of the platform's dependence on commercial LLMs.
• Discussion of alternative approaches for LLM implementation in transcriptomics.

Main Results:

• The AICellType platform demonstrates practical utility in cell type annotation.
• Continued reliance on proprietary LLMs like Claude 3.5 Sonnet poses challenges for transparency and reproducibility.
• Open-source LLMs, multimodal models, and local deployment offer potential improvements.

Conclusions:

• While AICellType is a valuable contribution, its dependence on commercial LLMs limits its potential.
• Future translational bioinformatics requires more emphasis on open-source, reproducible, and privacy-preserving annotation systems.
• Adoption of open-source LLMs, multimodal models, and local deployment is crucial for sustainable medical annotation.