CycleGPT

This repo is using for “Exploring the macrocyclic chemical space for heuristic drug design with deep learning models”.

Environment

conda env create -f CycleGPT.yaml  
conda activate CycleGPT 

After testing many times on many different machines, we recommended to install pytorch separately to accommodate compatibility issues between installation packages.

The following pytorch versions are used and tested at code time:

pip3 install numpy --pre torch --force-reinstall --index-url https://download.pytorch.org/whl/nightly/cu117

Usage of CycleGPT

Data processing

cd data
Macro_processing.py --macro_path xxx.csv --augment 0
Macro_processing.py --macro_path macro_all.csv --augment 0

Macro_processing.py: split and pad the data, implement augmentation if necessary(augment: switch to number you want).

xxx.csv can switch to your own data, such as: macro_all.csv.

Training

python lion_macro_train.py --input_name=xxx --device=cuda:2 --init_from=scratch --batch_size=128 --max_iters=30
python lion_macro_train.py --input_name=macro_all --device=cuda:2 --init_from=scratch --batch_size=128 --max_iters=30

input_name: train dataset name, the same as processing data,such as: macro_all.

init_from: choose whether to train the network from scratch or use the pre-trained network for transfer learning training.

init_from:scratch--init a new model.
init_from:resume--load a pretrain model. This need assign "resume_checkpoint"such as: --resume_checkpoint=xxx(pretrain model). If resume, max_iters may need to adjust.

Sampling

HyperTemp sampling as example. You can switch other sampling methods through sampling_methods.py.

Pretrain model for generated macrocycles can acquire from https://huggingface.co/FengAlbert/CycleGPT

python lion_sample_HyperTemp_transform.py --resume_checkpoint=xxx --temperature=0.7 --device=cuda:2
python lion_sample_HyperTemp_transform.py --resume_checkpoint=./ckpt_22_cycle_gpt.pt --temperature=0.7 --device=cuda:2

A trained model must be assigned to sample. --resume_checkpoint=xxx/xxx/xxx.pt (model file path).

This will generated a pkl file which contain generated molecules. Employing following script to decode molecules.

python macro_gen_load_gpt.py --input_name xxx  --pkl_path xxxx.pkl
python macro_gen_load_gpt.py --input_name macro_all  --pkl_path xxxx.pkl

Then valuate generated molecules:

macro_pick.py: seperate the macrocycles, linear molecules and invalid molecules.
macro_set.py: drop duplicate molecules.
macro_novelty.py: pick the novel molecules.

Usage of CyclePred

If you only use pre-trained CyclePred for prediction, you can skip the data processing and the training process. We provide pre-trained models in the ckpt folder.

We used dgl(1.1.1+cuda) version during code development, and now dgl maintains version 2.x. If the GPU is not available, the user can change devide to the cpu or update the dgl dependency package.

Data processing

cd CyclePred
python Macfrag_data.py

Using MacFrag to constract the Heterogeneous graph and Seperate the dataset for training.

Training

python CyclePred_train.py ./config/CyclePred_train_Jak2.json

Json file can switch the hyperparameter and train data.

Predicting

python CyclePred_predict_IC50.py ./config/CyclePred_predict_Jak2_IC50.json

Json file can switch the predicted model and predict molecules data.

Json file can switch the hyperparameter and train data.

Predicting

python CyclePred_predict_IC50.py ./config/CyclePred_predict_Jak2_IC50.json

Json file can switch the predicted model and predict molecules data.