---
license: mit

tags:
- chemistry
- flow matching
- proteins
- generative modelling
- mars-fm
- biology
- molecular-dynamics
- pytorch
---

# MarS-FM (MD-CATH) Model Checkpoint

This repository contains the PyTorch model weights for **MarS-FM** (Markov Space Flow Matching) trained on the **MD-CATH** dataset. This model was introduced in the ICLR 2026 paper: [*MarS-FM: Generative Modeling of Molecular Dynamics via Markov State Models*](https://arxiv.org/abs/2509.24779).

## 🔗 Links and Resources
* **Official GitHub Repository:** [valence-labs/mars-fm](https://github.com/valence-labs/mars-fm)
* **Paper:** [arXiv:2509.24779](https://arxiv.org/abs/2509.24779)
* **Dataset:** [compsciencelab/mdCATH](https://huggingface.co/datasets/compsciencelab/mdCATH)

## ⚙️ Model Configuration
* **Dataset:** MD-CATH (`--mdcath`)
* **Temperature:** 450K (`--temp 450`)
* **MSM States (`--msm_num_states`):** 10
* **MSM Lag Time (`--msm_lagtime`):** 50

## 💻 How to Use
For full instructions on how to set up the environment, load this checkpoint, and run generation sampling, please follow the documentation in the **[official GitHub repository](https://github.com/valence-labs/mars-fm)**.

## 📝 Citation
If you use these weights, please cite the original authors:

```bibtex
@article{kapusniak2025mars,
  title={MarS-FM: Generative Modeling of Molecular Dynamics via Markov State Models},
  author={Kapu{\'s}niak, Kacper and Gabellini, Cristian and Bronstein, Michael and Tossou, Prudencio and Di Giovanni, Francesco},
  journal={arXiv preprint arXiv:2509.24779},
  year={2025}
}
```