Leveraging large language models for literature-driven prioritization of protein binding pockets

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Identifying protein binding pockets is crucial for small-molecule drug discovery.
• Current manual methods for pocket identification are time-consuming, error-prone, and inefficient.
• Manual curation of binding site data hinders high-throughput drug discovery.

Purpose of the Study:

• To develop a novel, automated approach for identifying and prioritizing protein binding pockets.
• To leverage large language models (LLMs) for extracting residue-level information from scientific literature.
• To improve the efficiency and accuracy of small-molecule drug discovery pipelines.

Main Methods:

• Combined geometric pocket detection (Fpocket) with LLMs for pocket identification and validation.
• Utilized LLMs with fine-tuned prompts to extract residue information from published experimental data.
• Developed and employed a curated benchmark dataset for training and evaluating LLM performance.

Main Results:

• Successfully validated candidate pockets against published experimental data using LLMs.
• Demonstrated LLM's capability in assessing paper relevance and extracting crucial pocket information.
• Established a novel, automated workflow for protein binding pocket identification.

Conclusions:

• The combined geometric and LLM approach offers a significant advancement in automating protein binding pocket identification.
• This method addresses the limitations of manual curation, accelerating drug discovery.
• The developed dataset and methodology are publicly available to facilitate further research.