FFLOM Source Code

• Due to the size limit, train data and the pretrained checkpoints can be download from FFLOM | Zenodo.

• Users can go to ./tutorials fold to find two .ipynb file for the examples of either linker design or R-group design.

1. Install Environment

• Please create the environment using env.yaml .
  • The main module used in this code is RDKit and torch, you can also install RDKit by conda install -c rdkit rdkit and install torch on the official website.

2. Preprocess Dataset

We provide zinc250k dataset / casf dataset / PDBbind dataset in ./dataset fold. For example:

• To preprocess zinc_test dataset for linker design task, input codes below in the cmd:

• python preprocess.py --data dataset/linker/zinc_test.txt --save_fold ./data_preprocessed/zinc_test_linker/ --name zinc_test --linker_design

• To preprocess zinc_test dataset for R-group design task, input codes below in the cmd:

• python preprocess.py --data dataset/r_group/zinc_test.txt --save_fold ./data_preprocessed/zinc_test_r/ --name zinc_test --r_design

  • The input dataset should be in the separated SMILES format: 'molecule linker fragments' or 'molecule r-group fragment'. If you only get full molecules (in example.xls), please first use:

  • python cut_your_own_dataset.py --xls_path example.xls --output_path ./dataset/example.txt --linker_design --n_cores 4 --linker_min 5 --min_path_length 2

  • or R-group case:

  • python cut_your_own_dataset.py --xls_path example.xls --output_path ./dataset/example.txt --r_design --n_cores 4 --fragment_min 5

3. Training

We provide pretrained checkpoints in ./good_ckpt fold. To train your own model, use codes like:

• python train.py --path ./data_preprocessed/zinc_train/ --batch_size 32 --warm_up --epochs 30 --name train --seed 2019 --all_save_prefix ./

• To transfer learning on a small dataset (for example, like case studies) using the ZINC250K pretrained checkpoint 306, using codes like:

• python train.py --path ./data_preprocessed/case/ --batch_size 32 --warm_up --epochs 30 --name xx --seed 2019 --all_save_prefix ./ --init_checkpoint ./good_ckpt/checkpoint306

  • checkpoints will be saved in ./save_pretrain fold

4. Generation

• Given a checkpoint, you can generate molecules (length of generated fragments equal to the given ones) using codes like:

• python generate.py --path ./data_preprocessed/zinc_test_linker/ --gen_out_path ./mols/zinc_test_linker.txt --seed 66666666 --init_checkpoint ./good_ckpt/checkpoint306 --gen_num 10 

• If you want to design the length of generated fragments by your own, set len_freedom_x and len_freedom_y to become the range of length [len_freedom_x, len_freedom_y]:

• python generate.py --path ./data_preprocessed/zinc_test_linker/ --gen_out_path ./mols/test.txt --seed 66666666 --init_checkpoint ./good_ckpt/checkpoint306 --len_freedom_x -1 --len_freedom_y 1 --gen_num 10 

5. Evaluation

If you want to evaluate the generated molecules, using codes for linker design like:

python evaluate.py --train_data dataset/linker/zinc_train.txt --gen_data ./mols/test.txt --linker_design

codes for R-group design:

python evaluate.py --train_data dataset/r_group/zinc_train.txt --gen_data ./mols/test.txt --r_design

6. Reconstruction

For recovery metrics, please check recon.py for linker case or recon_r.py for r-group case:

python recon.py --path ./data_preprocessed/zinc_test_linker/ --seed 66666666 --init_checkpoint ./good_ckpt/checkpoint306

python recon_r.py --path ./data_preprocessed/zinc_test_r/ --seed 66666666 --init_checkpoint ./good_ckpt/checkpoint335