|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- chemistry |
|
|
- biology |
|
|
--- |
|
|
|
|
|
This repository provides the **UBio-Protein26 5M** dataset release used in the UBio-MolFM project: |
|
|
**"UBio-MolFM: A Universal Molecular Foundation Model for Bio-Systems"** ([link](https://github.com/IQuestLab/UBio-MolFM/blob/e2formerv2/MolFM-technique-report.pdf)). |
|
|
|
|
|
UBio-Protein26 5M is a protein-focused subset of UBio-Mol26, including solvated tripeptides and residue-centered protein clusters. |
|
|
|
|
|
## Computational Details |
|
|
|
|
|
All labels are generated from DFT calculations for physically consistent force-field training. |
|
|
|
|
|
- Functional: **ωB97M-D3** , following the setup described in the report. |
|
|
- Basis sets: |
|
|
- **def2-SVP** subset (majority, efficient large-scale coverage) |
|
|
- **def2-TZVPD** subset (high-fidelity subset; in practice with mixed-basis treatment for H/metal atoms as described in the report) |
|
|
- System scale: up to approximately **1,200 atoms**. |
|
|
- Training set size: **5M** samples |
|
|
- def2-SVP: **4500K** |
|
|
- def2-TZVPD: **500K** |
|
|
- Test set size: **200K** samples |
|
|
- def2-SVP: **180K** |
|
|
- def2-TZVPD: **20K** |
|
|
- Total size: **5.2M** samples. |
|
|
|
|
|
Note: The release name "5M" refers to the rounded training scale (about 5.0M) plus a 0.2M test split. |
|
|
|
|
|
## Dataset Layout |
|
|
The data is split first by `train/test`, then by basis set: |
|
|
|
|
|
```text |
|
|
UBio-Protein26-5M/ |
|
|
├── train/ |
|
|
│ ├── def2-svp/ |
|
|
│ └── def2-tzvpd/ |
|
|
└── test/ |
|
|
├── def2-svp/ |
|
|
└── def2-tzvpd/ |
|
|
``` |
|
|
|
|
|
All shards are in LMDB format (`data_XXXXX.lmdb`). |
|
|
|
|
|
 |
|
|
|
|
|
*Figure: Structural distribution of UBio-Protein26 5M, consistent with the report resources section.* |
|
|
|
|
|
## How to Load |
|
|
|
|
|
```python |
|
|
import lmdb |
|
|
import pickle |
|
|
|
|
|
lmdb_path = "train/def2-svp/data_00001.lmdb" |
|
|
|
|
|
env = lmdb.open(lmdb_path, readonly=True, lock=False, subdir=False) |
|
|
samples = [] |
|
|
|
|
|
with env.begin() as txn: |
|
|
# Dataset length is stored explicitly in LMDB. |
|
|
n = int(txn.get("length".encode())) |
|
|
for i in range(n): |
|
|
raw = txn.get(f"{i}".encode()) |
|
|
if raw is not None: |
|
|
samples.append(pickle.loads(raw)) |
|
|
|
|
|
env.close() |
|
|
|
|
|
if samples: |
|
|
print(samples[0].keys()) |
|
|
``` |
|
|
|
|
|
To compute total sample size for a split, sum `length` across all LMDB shards in that split/basis directory. |
|
|
|
|
|
## Data Fields |
|
|
|
|
|
This processed release keeps only the following fields: |
|
|
|
|
|
- `charge` |
|
|
- `multiplicity` |
|
|
- `pos` (shape `[N, 3]`, Angstrom) |
|
|
- `atomic_numbers` (shape `[N]`) |
|
|
- `energy` (unit: `kcal/mol`) |
|
|
- `forces` (shape `[N, 3]`, unit: `kcal/mol/Å`) |
|
|
|
|
|
Other auxiliary fields from upstream pipelines are intentionally removed. |
|
|
|
|
|
## Reference |
|
|
|
|
|
- Paper: [**UBio-MolFM: A Universal Molecular Foundation Model for Bio-Systems**](https://github.com/IQuestLab/UBio-MolFM/blob/e2formerv2/MolFM-technique-report.pdf) |
|
|
- Code (E2Former-V2): https://github.com/IQuestLab/UBio-MolFM/tree/e2formerv2 |
|
|
|
