---
license: mit
library_name: pytorch
tags:
  - chemistry
  - molecular-property-prediction
  - 3d-molecules
  - equivariant-gnn
  - molecular-embeddings
---

# MET-models

Checkpoint release for the MET project:

**Integrating equivariant architectures and charge supervision for data-efficient molecular property prediction**

## Links

- Paper DOI: https://doi.org/10.1039/D5ME00173K
- GitHub repository: https://github.com/mint258/MET
- Dataset repository: https://huggingface.co/datasets/Mint258/MET-dateset
- Intended model repository: https://huggingface.co/Mint258/MET-models

## Contents

This upload contains the pretrained and fine-tuned checkpoints used in the public MET repository.

### Pretrained Charge Models

- `pretrained_ckpt/best_model_dim8.pth`
- `pretrained_ckpt/best_model_dim16.pth`
- `pretrained_ckpt/best_model_dim32.pth`
- `pretrained_ckpt/best_model_dim64.pth`
- `pretrained_ckpt/best_model_dim128.pth`
- `pretrained_ckpt/best_model_dim256.pth`

### Fine-Tuned QM7 Models

- `fine-tuned_ckpt/qm7/qm7_data100.pth`
- `fine-tuned_ckpt/qm7/qm7_data300.pth`
- `fine-tuned_ckpt/qm7/qm7_data500.pth`
- `fine-tuned_ckpt/qm7/qm7_data800.pth`
- `fine-tuned_ckpt/qm7/qm7_data1000.pth`

### Fine-Tuned QM9 Dipole Models

- `fine-tuned_ckpt/dipole/dipole_data1000.pth`
- `fine-tuned_ckpt/dipole/dipole_data5000.pth`
- `fine-tuned_ckpt/dipole/dipole_data20000.pth`
- `fine-tuned_ckpt/dipole/dipole_data100000.pth`

## Recommended Local Layout

To keep the existing command examples in the MET GitHub repository unchanged, download these files into the same relative layout:

```text
pretrained_ckpt/
  best_model_dim8.pth
  best_model_dim16.pth
  best_model_dim32.pth
  best_model_dim64.pth
  best_model_dim128.pth
  best_model_dim256.pth
fine-tuned_ckpt/
  qm7/
    qm7_data100.pth
    qm7_data300.pth
    qm7_data500.pth
    qm7_data800.pth
    qm7_data1000.pth
  dipole/
    dipole_data1000.pth
    dipole_data5000.pth
    dipole_data20000.pth
    dipole_data100000.pth
```

## Checkpoint Notes

- `best_model_dim128.pth` is the main pretrained checkpoint used in the paper-facing examples and regression checks.
- The QM7 checkpoints correspond to atomization energy fine-tuning at different low-data subset sizes.
- The QM9 dipole checkpoints correspond to dipole-moment fine-tuning at different low-data subset sizes.

## Usage

Example charge evaluation:

```bash
python pretrain/charge_predict.py \
  --checkpoint_path pretrained_ckpt/best_model_dim128.pth \
  --test_data_root data/QM9/test_database.lmdb
```

Example QM7 evaluation:

```bash
python fine-tune/property_predict.py \
  --checkpoint fine-tuned_ckpt/qm7/qm7_data300.pth \
  --test_data_root data/QM7/test_database
```

## Notes

- This folder is prepared for manual upload to a Hugging Face Model repository.
- The checkpoints are distributed separately from the GitHub code repository to keep the source repository lightweight.