| --- |
| license: mit |
| --- |
| |
| # Data for FlashAffinity |
|
|
| This repository hosts the datasets and pre-trained model checkpoints for the paper: |
| **"FlashBind: Towards Accurate and Efficient Structure-based Virtual Screening"**. |
|
|
| The main code repository is available at: [🔗 GitHub: AIDD-Lab/FlashBind](https://github.com/AIDD-Lab/FlashBind). |
|
|
| ## Note |
|
|
| * **Enzyme Dataset (ESIBank):** For the enzyme dataset (`ESIBank.tar.zst`), we have not uploaded the pre-computed **ESM3 representations** due to their massive storage size. Users can generate these embeddings locally using the scripts provided in our [GitHub repository](https://github.com/AIDD-Lab/FlashAffinity). |
| * **Affinity Dataset (SAIR):** We utilize the **[SAIR Dataset](https://huggingface.co/datasets/SandboxAQ/SAIR)** for binding affinity values. We provide a complete preprocessing pipeline to handle this data. To prepare the dataset, please run the following scripts in order: |
| 1. `renumber.py` |
| 2. `structure_lmdb.py` |
| 3. `filter.py` |
| |
| *Please ensure you update the input parameters within each script before execution.* **Warning:** Due to the massive scale of the data, the preprocessing workflow takes approximately **12 hours (half a day)** and requires roughly **4–5 TB** of storage space. |
| |
| ## Citation |
|
|
| If you find FlashAffinity useful in your research, please cite our work: |
|
|
| ```bibtex |
| @article{jiang2025flashbind, |
| title={FlashBind: Towards Accurate and Efficient Structure-based Virtual Screening}, |
| author={Jiang, Songlin and Chen, Yifan and Krishnan, Aarti and Zhang, Yu and Jin, Wengong}, |
| journal={bioRxiv}, |
| } |
