Back

Assessing the Performance of LLMs in Multimodal Information Extraction for Biological Research: A Case Study on LLPS

Chin, K. Y.; Fujii, S.; Ishida, S.; Terayama, K.

2025-10-30 bioinformatics
10.1101/2025.10.29.685285 bioRxiv
Show abstract

Advances in experimental techniques have expanded the volume of biological data. This has increased the demand for structured information extraction from papers, with large language models (LLMs) considered promising. However, challenges remain, including limited validation in biology and unclear applicability to multimodal tasks that integrate text with domain-specific figures, such as microscopic images and scatter plots. Here, we developed a multimodal LLM (MLLM)-based workflow to extract the experimental conditions and phase status from the text and figures of experimental papers on liquid-liquid phase separation and validated the effect of various inputs, prompts, and MLLM types. As a result, the Gemini 2.5 Pro-based extraction achieved an F1-score of 0.847 by processing each figure as a processing unit and inputting domain-specific prompts reflecting manual extraction guidance. This study demonstrates the potential and limitations of MLLMs for extracting biological information and provides insights for advancing multimodal approaches in biology.

Matching journals

The top 7 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.