AI agents automate protein simulations, cutting setup time and human error

Molecular dynamics simulations are a cornerstone of structural biology, revealing how proteins fold, move, and function. Yet the process of setting up these computational experiments—preparing inputs, running simulations, analyzing outputs, and fixing problems—remains tedious and error-prone.

Researchers have now developed an automated framework that delegates these tasks to AI agents. The system, called NAMD-Agent, combines large language models with Python scripting and molecular simulation tools to handle the full pipeline without constant human oversight.

How it works#

The framework integrates LLM agents with Python and HTMD (a molecular simulation toolkit) to automatically generate inputs compatible with NAMD3 and CHARMM—standard software for molecular dynamics. Once a simulation is running, the agent monitors progress, analyzes results, and diagnoses and fixes failures that might otherwise halt computation.

Testing across diverse systems#

The researchers validated NAMD-Agent on five different types of biomolecular scenarios: a protein-DNA complex (the p53 DNA-binding domain), a membrane-embedded receptor (M2 muscarinic receptor), a series of ligand-bound proteins (five TYK2 inhibitors), a soluble protein (ubiquitin), and a protein-protein pair (barnase-barstar). For the DNA-bound system, the automated results matched those from an independently published benchmark study.

To separate the framework's capability from the quirks of any single AI model, the team repeated the protein-DNA experiment three times using different LLM orchestrators: Claude Opus-4.8, GPT-5.6 Sol, and Nemotron 3 Ultra. All three completed the full workflow, though they differed notably in accuracy, computational cost, and token usage—some variants consumed two orders of magnitude more tokens than others.

Why it matters#

This approach offers several practical advantages: it cuts setup overhead, reduces manual mistakes, enables parallel processing of many different protein systems, and provides a flexible foundation for scaling LLM-driven automation in structural biology. The pipeline can also call post-processing tools to refine results, creating a largely hands-free workflow from start to finish.