DiffPharm: Pharmacophore model guided 3D molecular generation through diffusion model

📦 Conda Environment Dependencies

Optimized Dependency List

• cudatoolkit == 11.8.0
• pytorch == 2.0.1
• rdkit == 2023.03.2
• scipy == 1.11.1
• hydra-core ==1.3.2
• imageio==2.31.1
• matplotlib==3.7.0
• numpy == 1.25.0
• omegaconf == 2.3.0
• pandas == 2.0.2
• Pillow==9.5.0
• pytorch_lightning == 2.0.6
• scikit_learn == 1.2.2
• setuptools==68.0.0
• torch_geometric == 2.3.1
• torchmetrics == 0.11.4
• tqdm == 4.65.0
• wandb == 0.15.4

You can create the environment with the following dependencies:

conda create -n diffpharm python=3.9 rdkit=2023.03.2
conda activate diffpharm


# Core dependencies
conda install -c "nvidia/label/cuda-11.8.0" cuda

pip3 install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

📂 Datasets

We use the same datasets as the MiDi model.

Download and place them under ./data/geom/raw/:

• Train: Download
• Validation: Download
• Test: Download

You can also use other datasets you want to use, such as CHEMBL, ZINC, or others.

---

🏋️ Training

We use MiDi's checkpoint geom-with-h-adaptive model as the pre-trained model. (We have also trained a new version of MiDi, link is here.)

Place it in:

./checkpoints/pre-trained/

Then run training, when we use the linker-free pharmacophore strategy. (removeHs=True)

cd ./midi
python main.py +experiment=linker_free_training

When we use the linker-align pharmacophore strategy. (removeHs=False)

cd ./midi
python main.py +experiment=linker_align_training

---

🧪 Testing

You can use:

• A model you trained in the previous step
• Or our trained model: Download

Place it in:

./checkpoints/diffpharm/

Then run:

cd ./midi
python main.py +experiment=linker_free_training general.test_only='ABS_PATH'
or
python main.py +experiment=linker_align_training general.test_only='ABS_PATH'

Replace ABS_PATH with the absolute path of the model checkpoint.

---

🧬 Sampling Example

1. Sampling via a Dedicated DiffPharm Script

This method is recommended for more lightweight or customized sampling scenarios. It directly invokes the dedicated sampling script diffpharm_sampling.py.py.

cd midi
python diffpharm_sampling.py +experiment=sampling

Make sure to set the correct statistics_path in your sampling.yaml.

🌍 Inpainting (Sampling under Dual Control: Substructure Fixing + Pharmacophore Guidance)

Inpainting allows you to design novel molecules around fixed substructures and pharmacophore-constraint, supporting tasks such as scaffold hopping, fragment linking, and fragment elaboration. You can either use your own trained model from the training step, or download our pre-trained model from the following link:

You can either use your own trained model from the training step, or download our pre-trained model:

📁 Pretrained Model on Google Drive

After downloading, place the model files into the ./checkpoints/diffpharm/ directory.

---

⚛️ How It Works

The inpainting workflow could refer to /inpainting_use/README.md