Design and Evaluation of an LLM-Based Knowledge Graph System for Photosynthesis Research

The exponential growth of scientific literature makes it increasingly difficult for researchers to organize and analyze large volumes of information, particularly in complex fields such as photosynthesis. Here, we present an automated knowledge graph construction system powered by large language models (LLMs) and validated on 5,000 photosynthesis research articles. A two-stage evaluation of eight LLMs showed marked performance differences: Gemini-2.5-flash achieved the highest F1-score (71.1%) and recall (78.0%), while GPT-4.1-nano demonstrated superior efficiency ($1.70/5k papers, 3× faster than Gemini-2.5-flash), optimizing real-time scalability. The system generated a large-scale network of 120,000 nodes and 540,000 relationships, enabling interactive exploration via a D3.js-based web platform. Domain coverage analysis revealed concentration in Plant Science (87.9%), Genetics (62.5%), and Molecular Biology (56.8%), with publication activity rising from 1995 to 2024 and peaking at 1,129 papers (2015–2019). Network analysis identified strong links among core fields but weak connections between Systems Biology and Circadian Rhythm, highlighting research gaps. The framework supports both beginner-friendly and advanced modes. While scalability and transferability remain challenges, this study demonstrates the potential of combining LLMs with knowledge graphs to accelerate discovery across the life sciences.