| --- |
| license: cc-by-nc-4.0 |
| tags: |
| - molecular-dynamics |
| - biomolecular-trajectories |
| - diffusion |
| - protein |
| - protein-ligand |
| - alphafold3 |
| library_name: protenix |
| --- |
| |
| # BioKinema: Physically Grounded Generative Modeling of All-Atom Biomolecular Dynamics |
|
|
| [](https://www.biorxiv.org/content/10.64898/2026.02.15.705956v1) |
| [](https://github.com/IDEA-XL/BioKinema/blob/main/LICENSE) |
| [](https://github.com/IDEA-XL/BioKinema/blob/main/DATA_LICENSE) |
| [](https://github.com/IDEA-XL/BioKinema) |
|
|
| ## Introduction |
|
|
| **BioKinema** is a physically grounded generative model that predicts continuous-time, all-atom biomolecular trajectories at a fraction of the cost of traditional molecular dynamics (MD) simulations. It is built on top of [Protenix](https://github.com/bytedance/Protenix) (ByteDance's AlphaFold 3 reproduction) and extends it with a temporal-attention mechanism derived from Langevin dynamics, so a single model can roll out MD-like trajectories at arbitrary, possibly non-uniform frame intervals. |
|
|
| The temporal-attention bias follows a stretched-exponential decay `B_ij = -λ |t_i - t_j|^β`, where `λ` is a per-head learnable decay (ALiBi-initialised) and `β` is a fixed time-scaling exponent selected per model variant. |
|
|
| > This HuggingFace repository hosts the released **weights** and processed **data**. For installation, inference, training, the data pipeline, and the manuscript benchmark code, see the [BioKinema GitHub repository](https://github.com/IDEA-XL/BioKinema). |
|
|
| ## Repository Contents |
|
|
| | File | Description | Size | |
| |------|-------------|------| |
| | `BioKinema_atlas+misato+mdposit_sqrt.pt` | **`sqrt`** checkpoint (EMA). For protein–ligand complexes and **short-time** MD. Trained on Atlas + MISATO + MDposit with `β = 0.5`. | ~3.9 GB | |
| | `BioKinema_CATH+octapeptide_beta0.25.pt` | **`beta=0.25`** checkpoint (EMA). For **long-time, single-chain** protein MD. Trained on MSR (CATH / MegaSim / octapeptides) with `β = 0.25`; adds a TICA-dynamics loss. | ~3.9 GB | |
| | `biokinema_codec_bundle.tar` | Processed MISATO / MDposit / unbinding data in a lossless compressed codec (one template bioassembly per trajectory + a stacked-coordinate array). Used by `sqrt` training. | ~41 GB | |
|
|
| ### Which Checkpoint to Use |
|
|
| - **Complexes / protein–ligand, or short MD** → `BioKinema_atlas+misato+mdposit_sqrt.pt` (run with `β = 0.5`). |
| - **Long single-chain protein MD / kinetics** → `BioKinema_CATH+octapeptide_beta0.25.pt` (run with `β = 0.25`). |
|
|
| The exponent `β` **must match the checkpoint** at inference time (pass it via `--beta`). |
|
|
| ## Usage |
|
|
| Clone the [BioKinema repository](https://github.com/IDEA-XL/BioKinema), install the environment, then run inference: |
|
|
| ```bash |
| bash inference.sh \ |
| --checkpoint_path ./checkpoints/BioKinema_atlas+misato+mdposit_sqrt.pt \ |
| --dump_dir ./output \ |
| --input_file ./experiments/atlas_benchmark/init_frames/7lp1_A_R1_0.cif \ |
| --beta 0.5 |
| ``` |
|
|
| DDP checkpoints (with a `module.` prefix) are handled automatically by the inference runner. |
|
|
| ### Codec Bundle |
|
|
| ```bash |
| tar -xf biokinema_codec_bundle.tar -C $BIOKINEMA_UNBINDING_ROOT |
| # -> $BIOKINEMA_UNBINDING_ROOT/{misato_codec,mdposit_codec,unbinding_codec} |
| ``` |
|
|
| Point each dataset's `bioassembly_dict_dir` at the corresponding `*_codec/` directory; the data loader auto-detects and decompresses on the fly. |
|
|
| ## Training Data |
|
|
| - **Atlas** — public ATLAS MD database (preprocessing scripts shipped in the code repository). |
| - **MSR (CATH / Octapeptides / MegaSim)** — Zenodo: `10.5281/zenodo.15629740`, `10.5281/zenodo.15641199`, `10.5281/zenodo.15641184`. |
| - **MISATO / MDposit / unbinding** — released here as the compressed codec bundle above. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{feng2026physically, |
| title={Physically Grounded Generative Modeling of All-Atom Biomolecular Dynamics}, |
| author={Feng, Bin and Zhang, Jiying and Zhang, Xinni and Zhang, Ming and Barth, Patrick and Liu, Zijing and Li, Yu}, |
| journal={bioRxiv}, |
| pages={2026--02}, |
| year={2026}, |
| publisher={Cold Spring Harbor Laboratory} |
| } |
| ``` |
|
|
| ## Acknowledgements |
|
|
| This project was built based on [Protenix](https://github.com/bytedance/Protenix), an open-source biomolecular structure prediction framework developed by ByteDance. |
|
|
| ## Contact |
|
|
| For questions or collaborations, please open an issue or contact us at [fengbin@idea.edu.cn](mailto:fengbin@idea.edu.cn). |
|
|
