Survey on large language model annotation of cellular senescence from figures in review articles.

This study evaluated large language models (LLMs), particularly the GPT-4 with vision (GPT-4 V) and GPT-4 Turbo, for annotating biomedical figures, focusing on cellular senescence. We assessed the ability of LLMs to categorize and annotate complex biomedical images to enhance their accuracy and efficiency. Our experiments employed prompt engineering with figures from review articles, achieving more than 70% accuracy for label extraction and approximately 80% accuracy for node-type classification. Challenges were noted in the correct annotation of the relationship between directionality and inhibitory processes, which were exacerbated as the number of nodes increased. Using figure legends was a more precise identification of sources and targets than using captions, but sometimes lacked pathway details. This study underscores the potential of LLMs in decoding biological mechanisms from text and outlines avenues for improving inhibitory relationship representations in biomedical informatics.