fengb commited on
Commit
ae43520
·
verified ·
1 Parent(s): 2603bea

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +91 -3
README.md CHANGED
@@ -1,3 +1,91 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ tags:
4
+ - molecular-dynamics
5
+ - biomolecular-trajectories
6
+ - diffusion
7
+ - protein
8
+ - protein-ligand
9
+ - alphafold3
10
+ library_name: protenix
11
+ ---
12
+
13
+ # BioKinema: Physically Grounded Generative Modeling of All-Atom Biomolecular Dynamics
14
+
15
+ [![Paper](https://img.shields.io/badge/Paper-bioRxiv-green)](https://www.biorxiv.org/content/10.64898/2026.02.15.705956v1)
16
+ [![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green?style=flat-square)](https://github.com/IDEA-XL/BioKinema/blob/main/LICENSE)
17
+ [![Data License](https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-red?style=flat-square)](https://github.com/IDEA-XL/BioKinema/blob/main/DATA_LICENSE)
18
+ [![GitHub Link](https://img.shields.io/badge/GitHub-blue?style=flat-square&logo=github)](https://github.com/IDEA-XL/BioKinema)
19
+
20
+ ## Introduction
21
+
22
+ **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.
23
+
24
+ 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.
25
+
26
+ > 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).
27
+
28
+ ## Repository Contents
29
+
30
+ | File | Description | Size |
31
+ |------|-------------|------|
32
+ | `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 |
33
+ | `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 |
34
+ | `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 |
35
+
36
+ ### Which Checkpoint to Use
37
+
38
+ - **Complexes / protein–ligand, or short MD** → `BioKinema_atlas+misato+mdposit_sqrt.pt` (run with `β = 0.5`).
39
+ - **Long single-chain protein MD / kinetics** → `BioKinema_CATH+octapeptide_beta0.25.pt` (run with `β = 0.25`).
40
+
41
+ The exponent `β` **must match the checkpoint** at inference time (pass it via `--beta`).
42
+
43
+ ## Usage
44
+
45
+ Clone the [BioKinema repository](https://github.com/IDEA-XL/BioKinema), install the environment, then run inference:
46
+
47
+ ```bash
48
+ bash inference.sh \
49
+ --checkpoint_path ./checkpoints/BioKinema_atlas+misato+mdposit_sqrt.pt \
50
+ --dump_dir ./output \
51
+ --input_file ./experiments/atlas_benchmark/init_frames/7lp1_A_R1_0.cif \
52
+ --beta 0.5
53
+ ```
54
+
55
+ DDP checkpoints (with a `module.` prefix) are handled automatically by the inference runner.
56
+
57
+ ### Codec Bundle
58
+
59
+ ```bash
60
+ tar -xf biokinema_codec_bundle.tar -C $BIOKINEMA_UNBINDING_ROOT
61
+ # -> $BIOKINEMA_UNBINDING_ROOT/{misato_codec,mdposit_codec,unbinding_codec}
62
+ ```
63
+
64
+ Point each dataset's `bioassembly_dict_dir` at the corresponding `*_codec/` directory; the data loader auto-detects and decompresses on the fly.
65
+
66
+ ## Training Data
67
+
68
+ - **Atlas** — public ATLAS MD database (preprocessing scripts shipped in the code repository).
69
+ - **MSR (CATH / Octapeptides / MegaSim)** — Zenodo: `10.5281/zenodo.15629740`, `10.5281/zenodo.15641199`, `10.5281/zenodo.15641184`.
70
+ - **MISATO / MDposit / unbinding** — released here as the compressed codec bundle above.
71
+
72
+ ## Citation
73
+
74
+ ```bibtex
75
+ @article{feng2026physically,
76
+ title={Physically Grounded Generative Modeling of All-Atom Biomolecular Dynamics},
77
+ author={Feng, Bin and Zhang, Jiying and Zhang, Xinni and Zhang, Ming and Barth, Patrick and Liu, Zijing and Li, Yu},
78
+ journal={bioRxiv},
79
+ pages={2026--02},
80
+ year={2026},
81
+ publisher={Cold Spring Harbor Laboratory}
82
+ }
83
+ ```
84
+
85
+ ## Acknowledgements
86
+
87
+ This project was built based on [Protenix](https://github.com/bytedance/Protenix), an open-source biomolecular structure prediction framework developed by ByteDance.
88
+
89
+ ## Contact
90
+
91
+ For questions or collaborations, please open an issue or contact us at [fengbin@idea.edu.cn](mailto:fengbin@idea.edu.cn).