MDCrow

Overview

MDCrow connects language models to tools for molecular dynamics rather than serving as a simulation engine itself. Built using Langchain, it is intended to coordinate tasks that would otherwise require separate preparation, simulation and analysis steps. The repository README identifies OpenMM as a particular execution target, while the research paper describes tools for file processing, simulation setup, analysis and retrieval from literature and databases.

The documented interface accepts a natural-language task through an MDCrow instance configured with an LLM model. The README illustrates a request containing a protein identifier, temperature, simulation duration and an instruction to calculate RMSD over time. This establishes an example workflow from a textual specification to simulation execution and a requested analysis; the supplied excerpts do not specify output file formats or a complete inventory of supported simulation settings. OpenAI is the default provider, with TogetherAI, Fireworks and Anthropic also documented as supported options requiring their respective integrations.

MDCrow has available repository code and installation guidance, so it is not a paper-only agent. Nevertheless, it should be approached as a research workflow tool: the paper evaluates task completion and sensitivity to model choice and prompting, not a guarantee of scientific validity for arbitrary simulations. Use requires a configured software environment and an LLM-provider API key. The repository’s MIT license covers its code; it does not establish terms for external model services or other dependencies.

Key Features

• Natural-language task submission through the MDCrow Python interface, with a configurable LLM model.
• Tools for preparing and executing molecular dynamics simulations, particularly through OpenMM.
• File handling and processing as part of multi-step molecular dynamics workflows, as described in the paper.
• Simulation-output analysis, including a README example requesting RMSD over time.
• Literature and database information retrieval described in the research paper.
• Documented integrations with OpenAI, TogetherAI, Fireworks and Anthropic; alternative providers require additional integration packages.

Use Cases

• Intended evaluation: reproduce the README’s protein-simulation and RMSD example, then inspect whether the generated setup and analysis match the requested conditions.
• Intended evaluation: assess whether natural-language orchestration can assist with a laboratory’s existing file-preparation, OpenMM simulation and analysis workflow.
• Intended evaluation: compare task execution under different documented LLM providers or prompt formulations without assuming the paper’s findings transfer to a new environment.
• Intended evaluation: explore literature or database retrieval alongside simulation planning, checking retrieved information before using it in scientific decisions.

How to Use

1. Read the README and paper to distinguish the runnable toolset from the research evaluation. Select a small, inspectable task for an initial trial.
2. Obtain the code from the repository. Follow the README’s conda environment procedure using environment.yaml, then its Git-based pip installation instructions; do not assume compatibility beyond the supplied guidance.
3. Choose a documented LLM provider. Follow the README’s API-key configuration guidance and use .env.example as a local template. Install the stated integration package if using an alternative provider, and keep credentials private.
4. Configure an MDCrow instance with the chosen model and submit a natural-language task through its run method. The README provides a protein-simulation and RMSD example and a Together-specific model-prefix convention.
5. As an evaluation step, inspect the simulation setup and resulting analysis before drawing scientific conclusions. Record the model, prompt and environment, and consult the paper when interpreting model- or prompt-dependent behavior.