MDCrow: automating molecular dynamics workflows with large language models

Area of Science:

• Computational Biology
• Biophysics
• Artificial Intelligence in Science

Background:

• Molecular dynamics (MD) simulations are crucial for understanding biomolecular systems.
• Automating complex MD workflows remains a significant challenge in scientific research.
• Large language models (LLMs) show promise in automating intricate scientific tasks via agentic approaches.

Purpose of the Study:

• To introduce MDCrow, an agentic LLM assistant designed to automate protein-focused MD workflows.
• To evaluate the capabilities of MDCrow in handling diverse and complex MD simulation tasks.
• To assess the impact of task complexity and prompt style on LLM agent performance.

Main Methods:

• MDCrow utilizes a chain-of-thought approach, leveraging over 40 specialized tools.
• Tools cover file management, simulation setup, output analysis, and literature/database retrieval.
• Performance was evaluated across 25 common MD tasks with varying complexity levels.

Main Results:

• MDCrow demonstrated proficiency in automating a wide range of protein MD simulation tasks.
• GPT-4o achieved high performance across tasks with minimal variance, followed by Llama3-405b.
• Prompt style significantly impacted smaller models but not the top-performing LLMs.

Conclusions:

• Agentic LLM assistants like MDCrow can effectively automate complex molecular dynamics workflows.
• Advanced models show robustness in handling complex scientific tasks, indicating a path towards automated scientific discovery.
• The development of specialized tools enhances the utility of LLMs for specific scientific domains like biophysics.