Compound generator: A GPT-based drug generator applied within a specific domain.

Workflow

Directory Structure

├── docker-compose.yml
├── Dockerfile
├── image
│   └── workflow.png
├── README.md
├── requirements.txt
└── src
    ├── config.json
    ├── dataset
    │   ├── Inhibitor
    │   │   ├── testing.csv
    │   │   └── training.csv
    │   └── sglt2
    │       └── sglt2.csv
    ├── logs
    ├── machine_model.py
    ├── main.py
    ├── ML_logs
    ├── model_save_path_reward
    ├── model_save_path_zinc20
    │   └── checkpoint-4468000
    ├── output
    │   └── reward_epoch
    ├── reward.py
    ├── trainer.py
    ├── visualize.ipynb
    └── zinc20M_gpt2_tokenizer

17 directories, 36 files

Requirements

CompoundGPT currently supports Python > 3.10

• joblib
• NumPy
• pandas
• scikit-learn
• rdkit
• pytorch > 2.0.0

Installation

pip or conda

To install PyTorch, visit the PyTorch official website and follow the instructions to download and install the PyTorch version that corresponds to your CUDA version.

Requirements can be installed using pip or conda as

pip install -r requirements.txt

or

conda list -e > requirements.txt

Docker

If you want to install CompoundGPT using a Docker, we have provided an image for your use.

First check if nvidia-docker is installed for GPU support, if not, please visit the Nvidia website

Build image

docker build -t compound:py310-torch211-cuda121 .

execute container

docker compose up -d

enter container

docker exec -it compound bash

Getting Started

Preparing Your Dataset

Format the Data: Prepare your dataset in the specified CSV format and place it under the /src/dataset/ directory. The expected format for the CSV should be as follows:

smiles, label

For an example of the file format, refer to the sample provided at here:

Configuration

Update Config File: Once your dataset is prepared and placed in the correct directory, navigate to the config.json file. Update the path in the config file to point to your newly processed dataset.

"train_data_path": "./dataset/your_path"
"kinase_name": "Name"

Training the Expert System

python machine_model.py

after finish training expert system it will show you performance.

Model Evaluation:
Sensitivity (Sn): 1.000
Specificity (Sp): 0.997
Accuracy (Acc): 0.999
Matthews Correlation Coefficient (MCC): 0.997

Finetune LLM

After completing the initial training of your model, the next step is to fine-tune your LLM.

Execute the following command to start the fine-tuning process.

CUDA_VISIBLE_DEVICES=1 python reward.py

During the training process, the model checkpoints for each epoch are saved in a specific directory. Here is how the saving mechanism is set up:

• Model Checkpoints : /src/model_save_path_reward
• Training Outputs : /src/output/reward_epoch

You can generate or check the differences for each epoch through these two locations.