MDForge agent designs molecular dynamics pipelines with sparse feedback
MDForge is an LLM agent that automates the design of molecular dynamics pipelines using open-ended code generation guided by verbal reward from multi-agent debate among physics experts, achieving results competitive with human experts on SAMPL benchmarks and discovering a novel picomolar CB[7] binder confirmed by wet-lab NMR.
Score breakdown
MDForge demonstrates that an LLM agent can autonomously design MD pipelines competitive with human experts and discover a wet-lab-validated picomolar binder, showing that agentic code generation with sparse feedback can replace expert-driven pipeline design in a domain where trial-and-error is computationally prohibitive.
- 01MDForge is an LLM agent that automates molecular dynamics (MD) pipeline design via open-ended code generation.
- 02Unlike prior MD agents, MDForge does not orchestrate a predefined tool set — pipeline design is treated as open-ended code generation.
- 03Sparse simulator feedback is densified using an in-context multi-agent debate among physics experts, which provides verbal reward.
Molecular dynamics (MD) is the standard computational method for simulating atomistic molecular behavior from first-principle physics, but designing an MD pipeline for a new system demands deep expert knowledge and is too costly for trial-and-error even on a single molecule. MDForge addresses this bottleneck by framing pipeline design as open-ended code generation rather than the orchestration of a fixed tool set — a key distinction from prior MD agents. An LLM agent's behavior is reshaped online by verbal reward, and because simulator feedback is sparse, MDForge densifies that reward signal through an in-context multi-agent debate among physics experts.
Evaluated on three SAMPL host-guest binding free-energy benchmarks, MDForge automatically designs MD pipelines that are competitive with those produced by human experts.
Evaluated on three SAMPL host-guest binding free-energy benchmarks, MDForge automatically designs MD pipelines that are competitive with those produced by human experts. When the system was deployed on a library of previously unseen candidate guest molecules, its CB[7] pipeline identified a novel binder; wet-lab competition NMR subsequently confirmed this molecule as a high-affinity, picomolar CB[7] binder. The paper's data and code are publicly available at the project's GitHub repository.
Key facts
- 01MDForge is an LLM agent that automates molecular dynamics (MD) pipeline design via open-ended code generation.
- 02Unlike prior MD agents, MDForge does not orchestrate a predefined tool set — pipeline design is treated as open-ended code generation.
- 03Sparse simulator feedback is densified using an in-context multi-agent debate among physics experts, which provides verbal reward.
- 04MDForge was evaluated on three SAMPL host-guest binding free-energy benchmarks, achieving results competitive with human experts.
- 05Deployed on a library of unseen candidate guest molecules, the CB[7] pipeline discovered a novel binder.
- 06Wet-lab competition NMR confirmed the discovered molecule is a high-affinity, picomolar CB[7] binder.
- 07Data and code are publicly available on GitHub at https://github.com/Zehong-Wang/MDForge.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 12, 2026 · 10:05 UTC. How this works →