Using Large Language Model to Optimize Protein Purification: Insights from Protein Structure Literature Associated with Protein Data Bank

Obtaining pure and homogeneous protein samples is vital for protein biology studies, yet optimizing protein expression and purification methods can be time-consuming because of variations in factors like expression conditions, buffer components, and fusion tags. With over 81 000 Protein Data Bank (PDB)-associated articles as of October 2024, manual extraction of relevant methods is impractical. To streamline this process, an automated tool is developed by incorporating a large language model (LLM) to extract and classify key data from these articles. The information extraction accuracy is enhanced by a 2-step-LLM and a 3-step-prompt. The key findings include: 1) Tris buffer is used in 49.2% of cases, followed by 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES) and phosphate buffers. 2) Polyhistidine tags dominate at 82.5%, followed by glutathione S-transferase (GST) and maltose-binding protein (MBP) tags. 3) E. coli expression is done at 16-20 °C, with induction period favoring 12-16 h (69.0%) over 3-6 h (14.3%). The statistical analyses highlight the correlation between protein properties and purification strategies. This tool is validated through two case studies: method bias for membrane protein purification, and crosslinker/detergent preferences for Cryo-Electron Microscopy sample preparation. These findings provide a valuable resource for designing protein expression and purification experiments.