FDC-Diff：Fragment-based Dual Conditional Diffusion Framework for Molecular Generation

🧩 Dependencies

This project requires a conda environment for dependency management. Please use the provided environment file to install all required packages.

🔧 Setup via Conda

# Clone the repository (if not done already)
git clone https://github.com/CHT713/FDC-Diff.git
cd FDC-Diff

# Create the environment from the YAML file
conda env create -f FC.yaml(If you encounter download failures, we recommend that you manually download each package that failed individually.)

# Activate the environment
conda activate FC

📦 Dataset Preparation

We use the CrossDocked dataset and the reaction-based slicing method from LibINVENT to construct datasets.To prepare the datasets from scratch, follow the steps below:

1.Download the dataset archive crossdocked_pocket10.tar.gz and the split file split_by_name.pt.(https://drive.google.com/drive/folders/1CzwxmTpjbrt83z_wBzcQncq84OVDPurM)

You can also find the original CrossDocked dataset at:https://bits.csb.pitt.edu/files/crossdock2020/

2.Extract the TAR archive using the command:

tar -xzvf crossdocked_pocket10.tar.gz

3.Split raw PL‑complexes and convert SDF to SMILES:

python split_and_convert.py

4.Use the reaction-based slicing method in LibINVENT to slice the molecules into scaffolds and R-groups in Lib-INVENT-dataset and replace and replace example_configurations/supporting_files/filter_conditions.json in Lib-INVENT-dataset with filter_conditions.json in this directory of data.

5.Process datasets：Switch to the data/single directory，Save the results obtained from libinvent, then modify the train_sliced_file, test_sliced_file (the save path for libinvent results), processed_train_file, and processed_test_file (the save path for the final processed data structure) in the data/single/process_and_prepare file.

python -W ignore process_and_prepare.py

📦 Training

Change data_path: /data/data in configs/single_full.yml to your data processing path.

python train.py --config configs/single_full.yml

📦 Sampling

Switch to the models folder under the yuel_bond directory, and run the following commands to download the bond prediction model weights:

wget https://zenodo.org/records/15353365/files/geom_3d.ckpt -O models/geom_3d.ckpt
wget https://zenodo.org/records/15353365/files/geom_cdg.ckpt -O models/geom_cdg.ckpt

And then, You need to modify the path in the def yuel_bond function in tools/bond_restruction.py, replacing it with your file path. It is recommended to use absolute paths to avoid unnecessary errors.

Futhermore, modify the places involving paths in sample.py and update them to your path address.

python sample.py

Adapt to other datasets

If you want to train and test other datasets on our model：

1. Convert all ligands to SMILES format and write them into a file with the .smi file extension. For example:

CN(CCC(N)CC(=O)NC1CCC(N2C=CC(N)(O)NC2=O)OC1C(=O)O)C(=N)N
COc1cc(OC)c(S(=O)(=O)NCc2ccccc2N2CCCCC2)cc1NC(C)=O
Nc1ncnc2c1ncn2C1OC(COP(=O)(O)OP(N)(=O)O)C(O)C1O
Nc1cc(S(O)(O)O)c(N)c2c1C(=O)c1ccccc1C2=O

2. Use LibINVENT to split ligand data into scaffolds and R-groups. (remeber to modify the reaction_based_slicing.json file in the LIBINVENT code. Modify its input and output paths).
3. Save the results obtained from libinvent, then modify the train_sliced_file, test_sliced_file (the save path for libinvent results), processed_train_file, and processed_test_file (the save path for the final processed data structure) in the data/single/process_and_prepare file. To adapt to your input of model, you need to modify the output of process_and_prepare.py to include the following structure：

1）xxx.csv

The structure of the CSV file is as follows:

uuid	molecule_name	molecule	scaffold	rgroups	anchor	pocket_full_size	pocket_bb_size	molecule_size	scaffold_size	rgroup_size	protein_filename	fragment
0	__4aaw_A_rec_4ac3_r83_lig_min.pdb	COc1cc(OC)c(S(=O)(=O)NCc2ccccc2N2CCCCC2)cc1NC(C)=O	COc1cc(OC)c(S(=O)(=O)NCc2ccccc2)cc1NC(C)=O	C1CCNCC1	0	215	112	31	25	6	/data/crossdocked_pocket10/GLMU_STRPN_2_459_0/4aaw_A_rec_4ac3_r83_lig_tt_min_0_pocket10.pdb	c1ccccc1

2）scaf.sdf

3) fragment.sdf

4) pockets.pkl

5) rgroup.sdf

6) mol.sdf

(You can refer to our code for handling.) Then modify the preprocess function of datasets.py to replace the paths with the paths of the data you generated above.

4. if you still encounter issues, please check whether the absolute path of data is used.