| --- |
| license: cc-by-4.0 |
| task_categories: |
| - other |
| tags: |
| - protein-ligand |
| - structure-based-drug-design |
| - molecular-generation |
| - anewomni |
| pretty_name: AnewOmni Training Datasets (non-SIU) |
| --- |
| |
| # AnewOmni Training Datasets — non-SIU subset |
|
|
| Preprocessed, **training-ready** datasets for [AnewOmni](https://github.com/bytedance/AnewOmni) |
| ("Programming Biomolecular Interactions with an All-Atom Generative Model"), in the codebase's |
| memory-mapped format (`data.bin` + `index.txt` + split/cluster index files). |
|
|
| This bundle holds **7 of the 8** training sources — **everything except SIU** (SIU is ~156 GB and is |
| distributed separately). After extraction the `datasets/` tree matches the paths in |
| `configs/train_ldm.yaml` / `configs/train_vae.yaml` with **0 missing files**. |
|
|
| ## Contents |
|
|
| | Dataset | Modality | Entries | Splits | Original source | |
| |---|---|---:|---|---| |
| | PepBench | peptide | 6,105 | train 4,157 / valid 114 (LNR 93 test held out) | Zenodo 13373108 (PepGLAD) | |
| | ProtFrag | peptide (augmentation) | 70,498 | — | Zenodo 13373108 (PepGLAD) | |
| | SAbDab | antibody | 16,947 | train 9,473 / valid 400 | OPIG / UniMoMo | |
| | PDBbind v2020 | small molecule | 16,200 | refined 4,628 (+train/valid) + other-PL 11,572 | PDBbind v2020 | |
| | BioLiP2-nr | small molecule | 51,114 | pretrain / finetune (by resolution) | zhanggroup.org/BioLiP | |
| | CrossDocked2020 | small molecule | 100,081 | train 99,881 / valid 100 / test 100 | TargetDiff Google Drive | |
| | _SIU (separate upload)_ | small molecule | 4,289,980 | cluster-weighted | HuggingFace `bgao95/SIU` | |
|
|
| Bundle total: **~261k entries, 17.2 GB** (`anewomni_datasets_noSIU.tar`). |
|
|
| ## Usage |
|
|
| ```bash |
| # on your training server, from the AnewOmni repo root: |
| hf download Windsao/anewomni-train-data anewomni_datasets_noSIU.tar --repo-type dataset --local-dir . |
| tar -xf anewomni_datasets_noSIU.tar # -> ./datasets/... |
| rm anewomni_datasets_noSIU.tar |
| ``` |
|
|
| The extracted layout: |
|
|
| ``` |
| datasets/ |
| peptide/pepbench/processed/ (+ ../train.cluster) |
| peptide/ProtFrag/processed/ |
| antibody/SAbDab/processed/ |
| molecule_v2/PDBbind/processed/{refined-set,v2020-other-PL}/ |
| molecule_v2/biolip2_nr/processed/ |
| molecule_v2/CrossDocked/processed/ |
| ``` |
|
|
| ## Format |
|
|
| Each dataset is a **memory-mapped store**: |
| - `data.bin` — concatenated, per-entry **zlib-compressed** complexes (receptor + binder). |
| - `index.txt` — one line per entry: `id \t start \t end \t properties-json` (byte range `[start,end)` into `data.bin`). |
| - `*_index.txt` — train / valid / test subsets (an entry list). |
| - `*.cluster` — sequence / scaffold clusters used for size- and redundancy-aware sampling. |
|
|
| Loadable via the AnewOmni `MoleculeDataset` / `PeptideDataset` / `AntibodyDataset` classes (`mmap_dir=…`, `specify_index=…`, `cluster=…`). |
|
|
| ## Notes |
|
|
| - **Counts are from raw processing of each source.** The AnewOmni paper applies an additional |
| ligand-quality filtering pipeline (CCD exclusion list, cross-source dedup by PDB code, 40% |
| sequence-identity train/test cutoff) that is **not** applied here — so the small-molecule counts |
| above are somewhat higher than the paper's final training numbers (PDBbind 14,200 / BioLiP2 18,012 / |
| CrossDocked 85,938). PepBench (4,157) and SAbDab (9,473/400) splits match the paper exactly. |
| - Derived from publicly available datasets (mostly CC-BY); refer to each original source for its license. |
|
|
