LatentFrag

KO | EN

gitlite — search

#javascript #hacktoberfest #typescript #golang #machine-learning #deep-learning #linux #java #windows #awesome #vue #pytorch #mcp #open-source

(★ 31)

Structure-based fragment identification in latent space

LatentFrag 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# LatentFrag

Structure-based fragment identification in latent space.

This work proposes a protein-fragment encoder that is trained contrastively between protein surfaces and fragments. This latent representation is next used in a generative model based on flow matching where we sample fragments in this learned latent space and their centroids conditioned on the pocket surface.

## Installation

### Option A: Modern environment (recommended)

Tested with Python 3.10, 3.11, and 3.12 using PyTorch 2.1+ and CUDA 12.1+.

```bash
cd LatentFrag
bash install.sh              # creates 'latentfrag' env with Python 3.11
# or: bash install.sh myenv 3.12   # custom env name / Python version

conda activate latentfrag
python test_setup.py         # verify
```

<details>
<summary>Manual installation steps</summary>

```bash
conda create --name latentfrag python=3.11 -y
conda activate latentfrag

# PyTorch + geometric
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
pip install torch-geometric==2.3.1
pip install torch-scatter torch-cluster -f https://data.pyg.org/whl/torch-2.1.0+cu121.html

# Core scientific deps
pip install "numpy<2" "setuptools<81"
pip install pytorch-lightning==2.1.0 wandb biopython pdb-tools ProDy plyfile pyvtk
pip install rdkit-pypi pandas matplotlib jupyter   # use 'rdkit' instead of 'rdkit-pypi' for Python 3.12
pip install posebusters fcd useful-rdkit-utils

# C++ standard library fix (needed on most systems)
conda install -c conda-forge libstdcxx-ng -y
mkdir -p "$CONDA_PREFIX/etc/conda/activate.d" "$CONDA_PREFIX/etc/conda/deactivate.d"
echo 'export _OLD_LD_LIBRARY_PATH="${LD_LIBRARY_PATH:-}"' > "$CONDA_PREFIX/etc/conda/activate.d/libstdcxx.sh"
echo 'export LD_LIBRARY_PATH="$CONDA_PREFIX/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"' >> "$CONDA_PREFIX/etc/conda/activate.d/libstdcxx.sh"
echo 'export LD_LIBRARY_PATH="$_OLD_LD_LIBRARY_PATH"; unset _OLD_LD_LIBRARY_PATH' > "$CONDA_PREFIX/etc/conda/deactivate.d/libstdcxx.sh"
conda deactivate && conda activate latentfrag

# Install LatentFrag
cd LatentFrag
pip install -e .
```

</details>

### Option B: Legacy environment (Python 3.7)

The models were trained with this setup, but it might not be compatible with modern CUDA/PyTorch versions. Use this if you have specific dependency constraints or want to exactly replicate the original training environment.

<details>
<summary>Click to expand legacy install instructions</summary>

```bash
conda create --name latentfrag python=3.7 -y
conda activate latentfrag

conda install cmake==3.27.4 -y
conda install cudatoolkit=11.3 -y
conda install cudatoolkit-dev=11.7 -c conda-forge -y
conda install cudnn=8.2.1 -y

pip install --upgrade pip
pip cache purge

pip install open3d==0.9.0
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
pip install torch-geometric==2.3.1 -f https://data.pyg.org/whl/torch-1.12.1+cu113.html

pip install pandas matplotlib jupyter pytorch-lightning==1.8.0 wandb==0.13.4 rdkit-pypi==2022.9.5 biopython==1.78 pdb-tools==2.5.0 ProDy==2.4.0
pip install plyfile==0.7.2 pyvtk==0.5.18
pip install --no-index torch-scatter torch-cluster -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
conda install -c conda-forge imageio
pip install posebusters fcd useful-rdkit-utils

cd LatentFrag
pip install -e .
```

</details>

Test the installation:

```bash
python test_setup.py
```

### Install reduce

Reduce is needed to protonate the protein structures. Follow instructions from the [reduced repo](https://github.com/rlabduke/reduce) to install reduce.

### Install MSMS

MSMS is needed to compute protein surfaces. Follow instructions from the [provider website](https://ccsb.scripps.edu/msms/downloads/) to install MSMS.

For example for linux unpack the downloaded binary like this:

```bash
tar zxvf msms_i86_64Linux2_2.6.1.tar.gz
```

### Install PLIP

PLIP is needed to compute protein-ligand interactions. Find more information on the [PLIP documentation](https://plip-tool.biotec.tu-dresden.de/plip-web/plip/index).

```bash
git clone https://github.com/pharmai/plip.git
cd plip
python setup.py install
sudo apt install plip
# check if plipcmd works
plipcmd -h
```

## Data

Download the data from [the figshare](https://figshare.com/s/c28e5d7f9d318305614e). We provide files necessary for inference and model checkpoints for both the encoder and the flow matching model. In order to generate the data for training, please follow the instructions in the `README_DATA.md`.

## Usage

### Encoder

#### Encode fragments

To compute the latent encodings of fragments, a sdf file containing a single fragments, a folder of sdf files or a SMILES string can be provided. For example:

```bash
python src/latentfrag/encoder/predict_frag.py --sdf_file demo_data/1MVC_BM6.sdf --checkpoint /path/to/encoder.ckpt --out_dir /path/to/out_dir
```

#### Encode a protein surface

To compute the latent encodings of a protein surface, a pdb file is needed. The script will compute the protein surface using MSMS and then compute the latent encoding (save under the variable `desc`). Ideally the protein gets protonated using reduced before.

```bash
python src/latentfrag/encoder/predict_protein.py --pdb_file demo_data/1MVC_A.pdb --checkpoint /path/to/encoder.ckpt --out_dir /path/to/out_dir --msms_bin /path/to/msms.x86_64Linux2.2.6.1
```

#### Training encoder

To train the encoder, a example config file is provided in `configs/train_encoder_example.yml`. Please adapt the data paths accordingly.

```bash
python src/latentfrag/encoder/train.py --config configs/train_encoder_example.yml
```

#### Make own fragment library

To create your own fragment library based on our encoder to be used with LatentFrag, you can use the following script. It will compute the latent encodings of all fragments in a single `.sdf` file. If there are multiple conformers for the same fragment, the lowest energy conformer will be used.

```bash
python src/latentfrag/encoder/make_frag_library.py --sdf_path /path/to/your_fragments.sdf --ckpt /path/to/encoder.ckpt --out /path/to/out_dir
```

### Generative modeling

#### Fragment identification with Flow Matching

We provide a processed PDB file and the corresponding reference ligand for a sampling demo (folder `demo_data`). There are different ways to define how many fragment per sample to generate:

- by number of reference fragments (`--sample_gt_size`): providing a reference ligand is necessary
- by chosen number (`--num_nodes`)
- by histogram based on pocket size (`--size_histogram`)

The pocket is defined either by a predefined radius around the reference ligand (`--reference_ligand`) or by defining pocket residues (`--residue_ids`). The following script defines the pocket and the sampled nodes by the ligand. Sampled fragments are not docked and thus are placed with arbitrary orientation at the predicted centroid. Please adjust the paths in the bash script accordingly.

```bash
cd LatentFrag
conda activate latentfrag
configs/demo_sampling.sh
```

If you also want to dock the sampled fragments, add the argument `--dock_fragments` and define the path to the gnina executable with `--gnina_exec`. Furthermore, if you provide your own protein, please make sure it is protonated for best performance. Protonation can be carried out in the same script when removing `--skip_protoonation` in the example demo and providing the path to the reduce executable with `--reduce_exec`.

#### Training FM

To train the flow matching model, a example config file is provided in `configs/train_flow_matching_example.yml`. Please adapt the data paths accordingly.

```bash
python src/latentfrag/flow_matching/train.py --config configs/train_flow_matching_example.yml
```

LatentFrag

LatentFrag 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

LatentFrag

LatentFrag 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README