Skip to content

DynaMate: An Autonomous Agent for Protein-Ligand Molecular Dynamics Simulations

🕒 Published (v1): 2025-12-10 19:40 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

DynaMate is a modular multi-agent LLM framework that autonomously executes complete molecular dynamics (MD) simulation workflows for protein and protein–ligand systems, including MM/PB(GB)SA binding free energy calculations. It combines a planner, MD execution agent, and analyzer with dynamic tool use, RAG-based literature retrieval (PaperQA + web search), and iterative self-correction to handle runtime errors. Evaluated across 12 benchmark systems with 5 LLMs, it achieves 100% accuracy on standard protein–ligand complexes.

Problem

Setting up MD simulations—structure cleaning, protonation, parameterization, solvation, multi-stage equilibration—is error-prone, labor-intensive, and requires expert knowledge, especially for protein–ligand systems where ligand parameterization must match the biomolecular force field. Existing automation frameworks (CHAPERONg, PyAutoFEP, MDCrow, NAMD-Agent) are rigid, domain-specific, or restricted to protein-only systems without adaptive error recovery.

Method

DynaMate is a three-agent pipeline running within a sandboxed directory: 1. Planner agent: Parses natural-language input, retrieves PDB structures, and generates a context-aware ordered simulation plan via RAG (PaperQA corpus + web search). 2. MD agent: Executes the plan through a tool-invocation → validation → reflection loop (hard cap: 35 iterations). On failure, it parses error messages and regenerates/edits input files (topology, .mdp, index files) to self-correct. 3. Analyzer agent: Post-processes trajectories (RMSD, RMSF, radius of gyration, hydrogen bonds) and produces natural-language summaries.

The framework is built on LiteLLM (v1.79.3) for unified model routing, uses GROMACS 2023 for simulation, and AmberTools/GAFF2 for ligand parameterization. Chat history is periodically compressed to avoid context overload.

Key Contributions

  • First agentic framework to autonomously handle full protein–ligand MD workflows, including ligand parameterization (GAFF2/AMBER ff14SB) and complex topology construction.
  • Self-correcting behavior via iterative error-parse → file-edit loops, demonstrated on three structured case studies (position restraints, wrong atom names, multi-ligand systems).
  • RAG integration combining PaperQA (curated literature corpus) and live web search to justify parameter choices with citable sources.
  • MM/PB(GB)SA binding free energy calculation module (gmx_MMPBSA) integrated into the workflow post-simulation.
  • Benchmarked across 12 systems (5 protein–ligand, 5 protein-only, 2 intentionally erroneous) with 5 LLMs evaluated on accuracy and tool-call efficiency.

Results

  • 100% pipeline accuracy on all five standard protein–ligand benchmark systems (3HTB, 3PTB, 4GIH, 4W52, 5UEZ) across all five LLMs.
  • Claude 3.7 Sonnet achieved 100% accuracy on the BRD4_UNL intentional-error system (wrong chlorine atom name); GPT-4.1 mini also solved it; Llama 3.3 failed all three runs on error-correction cases.
  • No LLM completed the two-ligand system (5KB6_ADN); Claude agents provided more actionable error diagnostics than GPT/Llama.
  • MMPBSA ΔΔG values for 10 BRD4 BD1 inhibitors correlated with experimental IC50 at r=0.597 vs. r=0.385 for GNINA 1.3 docking scores.
  • Claude 3 Opus and GPT-4.1 showed highest tool-call efficiency (fewest calls above the minimum required); Claude 3.7 Sonnet used more calls to recover from errors but achieved higher accuracy; Llama 3.3 underperformed across all metrics.

Limitations

  • Cannot handle systems with multiple ligands; multi-file parallel modification exceeds current toolset.
  • No persistent memory across runs; each run starts from scratch.
  • Strong dependence on underlying LLM quality—Llama 3.3 substantially underperforms Claude/GPT-4.1 on tool use and error correction.
  • Non-standard PDB atom names, missing residues, and incomplete coordinates remain persistent failure sources.
  • Protonation state assignment (histidine, aspartate, glutamate, ligands) is case-specific and not guaranteed correct; GAFF2 parameters should be verified by the user.
  • MM/PB(GB)SA inherits known approximations: implicit solvent, neglected conformational entropy, limited to comparative ranking of similar complexes.
  • Requires external LLM API keys/credits; rate limits and security concerns apply.

Relevance to Harnesses / Meta-Harnesses

DynaMate is a concrete domain-specific multi-agent harness instantiation: it separates high-level reasoning (planner) from low-level execution (MD agent) and post-processing (analyzer), with a hard iteration cap, structured tool schemas loaded at runtime, and periodic context compression—all standard harness engineering concerns. Its self-correcting loop (invoke → validate → reflect → patch) is a direct example of the error-recovery pattern central to robust scientific meta-harnesses. The RAG integration (PaperQA + web search) as a grounding layer for parameter decisions illustrates how domain knowledge retrieval can be wired into a harness to reduce hallucination in tool-use-heavy workflows. For researchers building general-purpose scientific harnesses, DynaMate demonstrates how to scope tool APIs, manage context length, and trade off accuracy vs. efficiency across LLM backends.