Add DeMiR pretrained weights and model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: pytorch
|
| 4 |
+
tags:
|
| 5 |
+
- chemistry
|
| 6 |
+
- molecular-representation
|
| 7 |
+
- density-matrix
|
| 8 |
+
- equivariant
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# DeMiR: Density-Matrix-Informed Representations for Molecular Property Prediction
|
| 12 |
+
|
| 13 |
+
Pretrained weights for DeMiR, a molecular representation model pretrained by
|
| 14 |
+
predicting DFT electron density matrices. Code: https://github.com/dhyeonk-0008/DeMiR
|
| 15 |
+
|
| 16 |
+
## Files
|
| 17 |
+
|
| 18 |
+
| File | Description | SHA-256 |
|
| 19 |
+
| --- | --- | --- |
|
| 20 |
+
| `demir.pth` | PyTorch `state_dict` (6.4M parameters, 25.8 MB) | `dbaefe607b31c4cccd6e0ea0c63eef40191c8903f055ee12eb89c614a88ed815` |
|
| 21 |
+
|
| 22 |
+
This is the checkpoint used for all results in the paper (validation-loss minimum
|
| 23 |
+
of the joint QM9 + QMugs training run).
|
| 24 |
+
|
| 25 |
+
## Usage
|
| 26 |
+
|
| 27 |
+
```bash
|
| 28 |
+
git clone https://github.com/dhyeonk-0008/DeMiR && cd DeMiR
|
| 29 |
+
# install as in the repository README, then:
|
| 30 |
+
hf download DeonX/DeMiR demir.pth --local-dir checkpoints
|
| 31 |
+
|
| 32 |
+
python scripts/inference/predict_density.py \
|
| 33 |
+
--smiles "CCO" \
|
| 34 |
+
--checkpoint checkpoints/demir.pth \
|
| 35 |
+
--output-npz ethanol.npz \
|
| 36 |
+
--device cpu
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
The architecture is inferred from the weights. See the repository README for
|
| 40 |
+
embedding extraction and property-prediction probes.
|
| 41 |
+
|
| 42 |
+
## Training data
|
| 43 |
+
|
| 44 |
+
B3LYP/cc-pVDZ density matrices computed with PySCF for QM9 molecules and a
|
| 45 |
+
representative subset of QMugs (adding S, P, Cl, Br and larger drug-like
|
| 46 |
+
molecules). Median real-space density NMAE on held-out molecules: 1.00% (QM9),
|
| 47 |
+
2.16% (QMugs).
|
| 48 |
+
|
| 49 |
+
## Limitations
|
| 50 |
+
|
| 51 |
+
- Neutral, closed-shell molecules in the cc-pVDZ basis only (H, C, N, O, F, P, S, Cl, Br).
|
| 52 |
+
- Bromine-containing molecules reconstruct poorly (NMAE > 8%).
|
| 53 |
+
- The eSCN backbone samples per-edge frames, so repeated forward passes differ
|
| 54 |
+
slightly (up to ~1e-2 in individual density-matrix elements) unless the random seed is fixed.
|
| 55 |
+
|
| 56 |
+
## License
|
| 57 |
+
|
| 58 |
+
MIT. The QM9 and QMugs source datasets keep their own licenses.
|
demir.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dbaefe607b31c4cccd6e0ea0c63eef40191c8903f055ee12eb89c614a88ed815
|
| 3 |
+
size 25827486
|