|
|
--- |
|
|
language: |
|
|
- en |
|
|
task_categories: |
|
|
- other |
|
|
tags: |
|
|
- chemistry |
|
|
- drug-discovery |
|
|
- molecule-generation |
|
|
- macrocyclic-peptide |
|
|
- 3d-generation |
|
|
- molecular-modeling |
|
|
--- |
|
|
|
|
|
# Unified All-Atom Molecule Generation with Neural Fields — MCPP Dataset |
|
|
|
|
|
This repository contains the Macrocyclic Peptide Pair (MCPP) dataset, curated for the paper [Unified all-atom molecule generation with neural fields](https://huggingface.co/papers/2511.15906). |
|
|
The code for the paper and model is available at: [https://github.com/prescient-design/funcbind/](https://github.com/prescient-design/funcbind/) |
|
|
|
|
|
We curated a dataset of **186,685 MCP–protein complexes** (`mcpp_dataset.tar.gz`) starting from **641 protein–MCP complexes** from the **[RCSB PDB](https://www.rcsb.org/)** using a **“mutate-then-relax”** strategy: |
|
|
|
|
|
## Dataset Generation Pipeline |
|
|
|
|
|
1. **Mutation:** |
|
|
MCPs were randomly mutated at **1 to 8 sites** using **213 distinct amino acids**. |
|
|
|
|
|
2. **Relaxation:** |
|
|
Mutated complexes were relaxed using **FastRelax in Rosetta**, which iteratively performs side-chain packing and all-atom minimization. |
|
|
|
|
|
3. **Selection:** |
|
|
The best complexes were chosen based on **lowest interface scores**. |
|
|
|
|
|
--- |
|
|
|
|
|
## Dataset Statistics |
|
|
- MCP lengths: **4–25 amino acids** (average 10) |
|
|
- **78%** of MCPs contain one or more **non-canonical amino acids** |
|
|
|
|
|
--- |
|
|
|
|
|
## Dataset Splits |
|
|
The dataset is split using a clustering-based approach. The **test set** covers **100 protein pockets**: |
|
|
|
|
|
| Split | File | |
|
|
|----------------|-----------------| |
|
|
| Training set | `train_data.pt` | |
|
|
| Validation set | `val_data.pt` | |
|
|
| Test set | `test_data.pt` |\ |
|
|
|
|
|
--- |
|
|
|
|
|
## Sample Usage |
|
|
|
|
|
This dataset provides preprocessed `.pt` files (`train_data.pt`, `val_data.pt`, `test_data.pt`) and the original `.tar.gz` file containing `.pdb` files. |
|
|
|
|
|
To use this dataset with the [FuncBind codebase](https://github.com/prescient-design/funcbind/): |
|
|
|
|
|
1. **Download and extract the original PDB files:** |
|
|
```bash |
|
|
tar -xvzf mcpp_dataset.tar.gz |
|
|
``` |
|
|
This will create a `mcpp_dataset/` directory containing the PDB files. |
|
|
|
|
|
2. **Place the preprocessed data:** |
|
|
Copy the `.pt` files into the extracted `mcpp_dataset/` directory. If you have cloned the FuncBind repository, the target path would be `funcbind/dataset/data/mcpp_dataset/`. |
|
|
```bash |
|
|
cp train_data.pt val_data.pt test_data.pt mcpp_dataset/ |
|
|
# Or if in FuncBind repo: |
|
|
# cp train_data.pt val_data.pt test_data.pt funcbind/dataset/data/mcpp_dataset/ |
|
|
``` |
|
|
|
|
|
3. **Alternatively, reprocess the data from scratch (within FuncBind repository):** |
|
|
After downloading and untarring `mcpp_dataset.tar.gz` into `funcbind/dataset/data/mcpp_dataset/`, ensure you have set up the FuncBind environment (see [GitHub repository](https://github.com/prescient-design/funcbind/quick-start)), then run: |
|
|
```bash |
|
|
cd funcbind/dataset |
|
|
python preprocess_mcp_pair.py |
|
|
``` |
|
|
|
|
|
4. **Sample Macrocyclic Peptides with FuncBind:** |
|
|
Once the data is prepared and FuncBind is installed (see [GitHub repository](https://github.com/prescient-design/funcbind/)), you can sample macrocyclic peptides. First, ensure you have downloaded the pre-trained checkpoints (e.g., `nf_unified` and `fb_unified`) from [Hugging Face](https://huggingface.co/mkirchmeyer/funcbind) and placed them in the appropriate `exps/` directories within the FuncBind repository. |
|
|
|
|
|
Then, from the FuncBind root directory, run: |
|
|
```bash |
|
|
python sample_fb.py --config-name sample_fb_mcpp |
|
|
``` |
