DanielYoung0520 commited on
Commit
d68d692
·
verified ·
1 Parent(s): 587c967

Align model card, setup commands, and repository contents with GitHub

Browse files
Files changed (2) hide show
  1. README.md +48 -41
  2. weights_manifest.json +1 -1
README.md CHANGED
@@ -14,38 +14,33 @@ inference: false
14
 
15
  <div align="center">
16
 
17
- <img src="images/oridynamic-logo.png" alt="OriDynamic" width="680">
18
 
19
  **All-atom conformational ensembles for biomolecular systems**
20
 
21
- v0 · Alpha inference weights
22
 
23
  [Code](https://github.com/JiahaoYoung0520/oridynamic_test_alpha) · [Installation](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/INSTALL.md) · [Examples](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/tree/main/examples) · [Model notes](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/model_card.md)
24
 
25
  </div>
26
 
27
- OriDynamic generates conformational ensembles from molecular sequences and composition, with an optional initial structure. A structural autoencoder establishes a latent representation, a flow module generates latent states that decode into representative conformations, and a refine module samples all-atom configurations conditioned on those proposals.
28
 
29
- The inference interface accepts protein monomers, protein–ligand complexes, and protein–protein assemblies. When an initial structure is omitted, the cofolding backbone predicts one before ensemble generation. The resulting ensemble is an unordered collection of conformations.
30
 
31
- ## Model files
32
 
33
- This repository provides two OriDynamic inference checkpoints for the [alpha code release](https://github.com/JiahaoYoung0520/oridynamic_test_alpha). The compatible code commit is recorded in [weights_manifest.json](weights_manifest.json). Download the Protenix v1.0.0 backbone separately from its official distribution.
34
-
35
- | File | Role | Size |
36
  | :--- | :--- | ---: |
37
- | `ae.ckpt` | Structural autoencoder | 0.81 GB |
38
- | `oridynamic.ckpt` | Latent flow module, coordinate refiner, and trunk adaptation (Joint200) | 0.95 GB |
39
-
40
- Joint200 is the checkpoint saved after **200 predicted-anchor adaptation updates from Joint11500**. The frozen codec, latent flow module, and refiner use the shared online LoRA trunk context. This requires the matching updated inference code; changing the weight file alone in the earlier code is insufficient.
41
 
42
- The OriDynamic download is approximately **1.76 GB**. The separately obtained official Protenix v1.0.0 backbone is approximately **1.48 GB**. The joint export includes the flow module; no separate flow checkpoint is needed. Filenames match the updated default path configuration; the loader verifies checkpoint identity independently of filenames. The AE and Joint checkpoint bytes are unchanged from the preceding release.
43
 
44
- These files contain inference tensors, required configuration, and latent normalization information. They do not contain optimizers or training datasets. File sizes and SHA-256 checksums are provided in [weights_manifest.json](weights_manifest.json). The loader checks the export schemas and the correspondence between the backbone, autoencoder, and Joint weights.
45
 
46
- ## Quick start
47
-
48
- The tested inference environment uses **Linux, Python 3.11, PyTorch 2.7.1 with CUDA 12.6, and an NVIDIA GPU**. Install the code and dependencies:
49
 
50
  ```bash
51
  git clone https://github.com/JiahaoYoung0520/oridynamic_test_alpha.git
@@ -56,18 +51,18 @@ conda activate oridynamic
56
 
57
  python -m pip install torch==2.7.1 --index-url https://download.pytorch.org/whl/cu126
58
  python -m pip install -e .
59
- python -m pip install -U huggingface_hub
60
  ```
61
 
62
- Download the two OriDynamic files and the official Protenix base into the code repository `weights/` directory. Downloads are public and do not require `hf auth login`.
63
 
64
  ```bash
 
65
  python -m pip install -U huggingface_hub
66
  hf download DanielYoung0520/oridynamic_v0 \
67
  ae.ckpt oridynamic.ckpt weights_manifest.json \
68
  --local-dir weights
69
 
70
- # Obtain the exact Protenix v1 base from its official distribution.
71
  curl -fL --retry 3 \
72
  https://protenix.tos-cn-beijing.volces.com/checkpoint/protenix_base_default_v1.0.0.pt \
73
  -o weights/protenix_base_default_v1.0.0.pt
@@ -76,17 +71,30 @@ printf '%s %s\n' \
76
  weights/protenix_base_default_v1.0.0.pt | sha256sum -c -
77
  ```
78
 
79
- Prepare the CCD resources following the [installation guide](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/INSTALL.md#2-model-weights-and-ccd), then set their location:
 
 
 
 
 
 
 
 
 
 
 
80
 
81
  ```bash
82
  export ORIDYNAMIC_PATHS="$PWD/configs/portable_paths.example.json"
83
- export ORIDYNAMIC_ASSETS="/absolute/path/to/assets"
84
  oridynamic doctor
85
  ```
86
 
87
- `ORIDYNAMIC_ASSETS` must contain `common/components.cif` and `common/components.cif.rdkit_mol.pkl`. CCD resources are separate from this weight bundle.
88
 
89
- Run the five examples provided with the code:
 
 
90
 
91
  ```bash
92
  oridynamic sample \
@@ -95,33 +103,32 @@ oridynamic sample \
95
  --out runs/five_cases
96
  ```
97
 
98
- The examples include **three monomers, one protein–ligand complex, and one protein–protein assembly**, with six precomputed protein MSAs. No MSA search tools or databases are needed for these examples. The command generates two final conformations per system. Use a new output directory for each run.
99
 
100
  ## Inputs and outputs
101
 
102
- | System | Input |
103
  | :--- | :--- |
104
- | Protein monomer | FASTA sequence and a matching A3M alignment |
105
- | Protein–ligand complex | JSON describing the protein chains, their MSAs, and ligand identities |
106
- | Protein–protein assembly | JSON describing the interacting chains and their MSAs |
 
 
 
 
107
 
108
- An initial PDB/mmCIF structure can be supplied for any supported system. For new sequences, provide matching MSAs or use the local ColabFold/MMseqs2 search option. See the [input guide](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/complex_inputs.md).
109
 
110
- Each target's output directory contains `anchor.cif`, `proposals.cif`, `ensemble.cif`, and run metadata. `ensemble.cif` contains the final generated conformations, with a total sample count equal to the number of seeds multiplied by `samples-per-seed`.
111
 
112
- ## Scope
113
 
114
- - These weights match the existing alpha inference implementation. They should not be mixed with checkpoints from other model branches or training runs.
115
- - Joint200 was adapted using labeled monomer and protein–ligand systems from the original test split. Results on these systems are not independent held-out evaluation and should not be equated with earlier Joint11500 benchmark results.
116
- - Protein–protein systems were excluded from this adaptation and remain outside its training scope, although the inference interface accepts them.
117
- - The exported tensors are verified against the source checkpoint. Full-size CPU loading and small-module sampling regressions check compatibility and the online codec route. An author-supplied H20 deployment report dated 2026-11-01 records Joint200 execution with five bundled MSA examples and ten conformations, plus twelve samples each for one monomer, PL, and PP system. Raw H20 outputs have not been independently reviewed in this release update. The report used an existing dependency image and does not validate a fresh public Dockerfile build. These checks do not establish ensemble quality.
118
- - The output is a time-independent ensemble. It does not provide a molecular-dynamics trajectory, physical transition times, or calibrated thermodynamic state populations.
119
- - GPU memory requirements depend on system size and sampling settings. Arbitrary-length inference on smaller GPUs has not been established.
120
 
121
- Seeds do not guarantee identical GPU coordinates. Changing sample chunk size can change intermediate stochastic noise assignment and produce different conformations. Keep sampling settings and the runtime environment fixed for comparisons; see the [reproducibility notes](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/model_card.md#reproducibility).
122
 
123
  ## Attribution and license
124
 
125
- The OriDynamic autoencoder and joint-model inference weights in this repository are released under the [Apache License 2.0](LICENSE). Copyright 2026 LGLab Frontier AI Group for the OriDynamic contributions and model adaptations.
126
 
127
- OriDynamic reuses the Protenix cofolding architecture and adapts its model parameters. The separately obtained Protenix v1 backbone retains its upstream Apache-2.0 license; a copy is provided in [LICENSE-PTX](LICENSE-PTX). Upstream copyright notices are preserved; see [NOTICE](NOTICE) and [source attribution](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/SOURCES.md). Third-party dependencies and external data resources retain their respective terms.
 
14
 
15
  <div align="center">
16
 
17
+ <img src="https://huggingface.co/DanielYoung0520/oridynamic_v0/resolve/main/images/oridynamic-logo.png" alt="OriDynamic" width="680">
18
 
19
  **All-atom conformational ensembles for biomolecular systems**
20
 
21
+ Alpha inference weights
22
 
23
  [Code](https://github.com/JiahaoYoung0520/oridynamic_test_alpha) · [Installation](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/INSTALL.md) · [Examples](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/tree/main/examples) · [Model notes](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/model_card.md)
24
 
25
  </div>
26
 
27
+ OriDynamic generates all-atom conformational ensembles for protein monomers, protein–ligand complexes, and protein–protein assemblies. Provide molecular sequences and composition, together with matching MSAs or a local MSA database. An initial structure is optional; when omitted, the cofolding backbone predicts one before ensemble generation.
28
 
29
+ ## Repository contents
30
 
31
+ This Hugging Face repository hosts **two OriDynamic inference checkpoints**, a weight manifest, the logo, and license files. Inference code, configurations, example inputs, and the six example MSAs are in the [GitHub code repository](https://github.com/JiahaoYoung0520/oridynamic_test_alpha). The Protenix base and CCD resources are downloaded separately from the official Protenix distribution.
32
 
33
+ | File | Contents | Size |
 
 
34
  | :--- | :--- | ---: |
35
+ | [`ae.ckpt`](https://huggingface.co/DanielYoung0520/oridynamic_v0/resolve/main/ae.ckpt) | Structural autoencoder | 0.81 GB |
36
+ | [`oridynamic.ckpt`](https://huggingface.co/DanielYoung0520/oridynamic_v0/resolve/main/oridynamic.ckpt) | Latent flow, coordinate refiner, trunk adaptation, and latent normalization | 0.95 GB |
37
+ | [`weights_manifest.json`](https://huggingface.co/DanielYoung0520/oridynamic_v0/blob/main/weights_manifest.json) | File sizes, SHA-256 hashes, compatible code commit, and external Protenix dependency | — |
 
38
 
39
+ The two OriDynamic checkpoints total approximately **1.76 GB**. The official Protenix base is approximately **1.48 GB** in addition. There is no separate flow checkpoint. Use the matching code and files identified in the manifest; filenames alone do not establish compatibility.
40
 
41
+ ## Installation and resources
42
 
43
+ Use **Linux, Python 3.11, PyTorch 2.7.1 with CUDA 12.6, and an NVIDIA GPU**. Start from the code checkout, not a clone of this weight repository:
 
 
44
 
45
  ```bash
46
  git clone https://github.com/JiahaoYoung0520/oridynamic_test_alpha.git
 
51
 
52
  python -m pip install torch==2.7.1 --index-url https://download.pytorch.org/whl/cu126
53
  python -m pip install -e .
54
+ oridynamic --help
55
  ```
56
 
57
+ Keep the environment activated and run all remaining commands from `oridynamic_test_alpha/`. Download the two OriDynamic checkpoints and the exact official Protenix v1.0.0 base:
58
 
59
  ```bash
60
+ mkdir -p weights assets/common
61
  python -m pip install -U huggingface_hub
62
  hf download DanielYoung0520/oridynamic_v0 \
63
  ae.ckpt oridynamic.ckpt weights_manifest.json \
64
  --local-dir weights
65
 
 
66
  curl -fL --retry 3 \
67
  https://protenix.tos-cn-beijing.volces.com/checkpoint/protenix_base_default_v1.0.0.pt \
68
  -o weights/protenix_base_default_v1.0.0.pt
 
71
  weights/protenix_base_default_v1.0.0.pt | sha256sum -c -
72
  ```
73
 
74
+ These downloads are public and do not require `hf auth login`. Download the two required CCD resources from the official source:
75
+
76
+ ```bash
77
+ curl -fL --retry 3 \
78
+ https://protenix.tos-cn-beijing.volces.com/common/components.cif \
79
+ -o assets/common/components.cif
80
+ curl -fL --retry 3 \
81
+ https://protenix.tos-cn-beijing.volces.com/common/components.cif.rdkit_mol.pkl \
82
+ -o assets/common/components.cif.rdkit_mol.pkl
83
+ ```
84
+
85
+ Configure the local paths:
86
 
87
  ```bash
88
  export ORIDYNAMIC_PATHS="$PWD/configs/portable_paths.example.json"
89
+ export ORIDYNAMIC_ASSETS="$PWD/assets"
90
  oridynamic doctor
91
  ```
92
 
93
+ The directory referenced by `ORIDYNAMIC_ASSETS` must contain `common/components.cif` and `common/components.cif.rdkit_mol.pkl`. The path configuration is supplied by the code repository and expects all three checkpoints under its `weights/` directory. `oridynamic doctor` checks installation and file paths; it does not load the models. Alternative locations, Docker, and local MSA setup are described in [INSTALL.md](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/INSTALL.md).
94
 
95
+ ## Run the bundled examples
96
+
97
+ From the code repository root:
98
 
99
  ```bash
100
  oridynamic sample \
 
103
  --out runs/five_cases
104
  ```
105
 
106
+ The example input contains **three monomers, one protein–ligand complex, and one protein–protein assembly**, with six precomputed protein MSAs. It requires no MSA search tools or databases. The command produces **two conformations per system, ten in total**. Final ensembles are saved under `runs/five_cases/targets/<name>/ensemble.cif`. Use a new output directory for each run.
107
 
108
  ## Inputs and outputs
109
 
110
+ | System | Molecular input |
111
  | :--- | :--- |
112
+ | Protein monomer | A protein sequence in FASTA or JSON |
113
+ | Protein–ligand complex | JSON containing protein sequences and CCD, SMILES, or ligand-file definitions |
114
+ | Protein–protein assembly | JSON containing the interacting protein chains |
115
+
116
+ Each protein entity needs a matching MSA, supplied as A3M or prepared by local search. The five bundled examples already include their MSAs. `--single-sequence` explicitly permits query-only input for smoke tests.
117
+
118
+ An optional PDB/mmCIF initial structure must match the input composition and contain the required heavy atoms. See the [input guide](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/complex_inputs.md) for chain IDs, ligand definitions, MSAs, and structure matching.
119
 
120
+ For `oridynamic sample`, each target is written to `<out>/targets/<name>/`, including `anchor.cif`, `proposals.cif`, `ensemble.cif`, and metadata. Batch summaries and timing are under `<out>/`. The final conformation count per target equals the number of seeds multiplied by `--samples-per-seed`; see the [output reference](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/batch_inference.md#outputs-and-timing) for exact file locations.
121
 
122
+ ## Model scope
123
 
124
+ The released joint model received additional adaptation on labeled monomer and protein–ligand systems from the original test split. Results on those systems are not independent held-out evaluation. Protein–protein systems were not included in that additional adaptation.
125
 
126
+ The code includes CPU regression tests, package checks, and strict checkpoint loading checks. An author-supplied H20 report records successful execution with these weights on the five bundled MSA examples. This is execution evidence, not an independent quality benchmark; the report used an existing dependency image rather than a fresh build of the public Dockerfile. Checkpoint history and the full validation scope are maintained in the [model notes](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/model_card.md#validation-and-scope).
 
 
 
 
 
127
 
128
+ Generated structures form an unordered ensemble, not an MD trajectory. Physical transition times and calibrated basin populations are not provided. GPU memory requirements depend on system size and sampling settings. Matching seeds do not guarantee identical GPU coordinates, and changing sample chunk size can change stochastic samples; see [reproducibility](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/model_card.md#reproducibility).
129
 
130
  ## Attribution and license
131
 
132
+ The OriDynamic weights are released under [Apache-2.0](https://huggingface.co/DanielYoung0520/oridynamic_v0/blob/main/LICENSE). Copyright 2026 LGLab Frontier AI Group for the OriDynamic contributions and model adaptations. The inference code is released under the same license in the GitHub repository.
133
 
134
+ OriDynamic reuses the Protenix cofolding architecture and adapts its parameters. The separately obtained Protenix v1 backbone retains its upstream license; a copy is provided in [LICENSE-PTX](https://huggingface.co/DanielYoung0520/oridynamic_v0/blob/main/LICENSE-PTX). See [NOTICE](https://huggingface.co/DanielYoung0520/oridynamic_v0/blob/main/NOTICE) and [source attribution](https://github.com/JiahaoYoung0520/oridynamic_test_alpha/blob/main/docs/SOURCES.md). Third-party resources retain their respective terms.
weights_manifest.json CHANGED
@@ -2,7 +2,7 @@
2
  "schema": "oridynamic_weights_bundle_v2",
3
  "release": "alpha",
4
  "code_repository": "https://github.com/JiahaoYoung0520/oridynamic_test_alpha",
5
- "compatible_code_commit": "0a0d3112cefc9fe712b98b78d49e1fb5b1c8818d",
6
  "weights": {
7
  "codec": {
8
  "file": "ae.ckpt",
 
2
  "schema": "oridynamic_weights_bundle_v2",
3
  "release": "alpha",
4
  "code_repository": "https://github.com/JiahaoYoung0520/oridynamic_test_alpha",
5
+ "compatible_code_commit": "bdcef5e2e435a41b369461b2b3f45214a4d3f8ec",
6
  "weights": {
7
  "codec": {
8
  "file": "ae.ckpt",