ADME Multitask Training and Fine-Tuning

This README provides an overview of the environment setup and directory/configuration file organization required for multitask Graph Neural Network (GNN) training on ADME property prediction and subsequent fine-tuning, as employed in the paper “Improving ADME Prediction with Multitask Graph Neural Networks and Assessing Explainability in Lead Optimization” by Ito S. et al.
By following this document, you will understand the essential requirements for data preparation and configuration placement. Detailed training and fine-tuning procedures are not covered here.

Contact

• Data Curator: Shoma Ito
• Repository Manager: Takuto Koyama (koyama.takuto.82j[at]st.kyoto-u.ac.jp)

Table of Contents

• ADME Multitask Training and Fine-Tuning
  • Contact
  • Table of Contents
  • Prerequisites
  • Directory Structure
    • Key subdirectories
  • Data Preparation
  • Training and Prediction Procedures
    • 1. Multitask Training
    • 2. Fine-tuning
    • 3. Prediction
    • Notes

Prerequisites

• kmol v1.1.7 (GitHub): the command-line tool for model training and evaluation

  Install required packages in a dedicated environment:

  git clone https://github.com/elix-tech/kmol.git
  cd kmol
  git checkout v1.1.7
  make create-env
  conda activate kmol

Directory Structure

ADMET_MTFT/
├── dataset/
│   └── all_train_data_2022/
│       ├── standardize-drug_list/
│       │   └── ADME/
│       │       └── all_data/
│       │           └── log_data/           # per-target log-transformed values
│       └── multitask.csv                   # combined multitask labels and values
└── configs/
    └── accuracy_drug/
        └── adme/
            ├── multitask/                  # config files for GNN multitask training
            └── finetuning_learning_rate/  # config files for GNN fine-tuning

Key subdirectories

• dataset/all_train_data_2022:

  • standardize-drug_list/ADME/all_data/log_data: per‑target log‑transformed values
  • .../multitask.csv: combined multitask labels and values

• configs/accuracy_drug/adme:

  • multitask/: config files for GNN multitask training
  • finetunig_learning_rate/: config files for GNN fine‑tuning

Data Preparation

Before training, ensure your data is properly formatted:

1. Individual target data: Place log-transformed values in dataset/all_train_data_2022/standardize-drug_list/ADME/all_data/log_data/
2. Multitask data: Combine all targets into dataset/all_train_data_2022/multitask.csv
3. Configuration: Adjust paths in config files to match your directory structure

For detailed data preprocessing steps (including log transformation, duplicate removal, and epsilon handling), please refer to: dataset/all_train_data_2022/standardize-drug_list/ADME/all_data/README.md

Training and Prediction Procedures

1. Multitask Training

To perform multitask GNN training on ADME properties:

# Activate the kmol environment
conda activate kmol

# Run multitask training using the configuration files
kmol train /path/to/ADME_MTFT/configs/accuracy_drug/adme/multitask/train/itr1/config.json

The multitask training will use the combined dataset (dataset/all_train_data_2022/standardize-drug_list/ADME/all_data/log_data/multitask.csv) to train a single GNN model on multiple ADME properties simultaneously.

Key configuration features for multitask training:

• Multitask loader configuration:

"loader": {
  "type": "multitask",
  "input_path": "multitask.csv",
  "task_column_name": "target_id",
  "max_num_tasks": 10
}

• Target-specific standardization: Each ADME property (target 0-9) has individual mean and standard deviation values for proper normalization:

"transformers": [
  {
    "type": "standardize",
    "target": 0,
    "mean": 1.5067472653927056,
    "std": 0.6985528512457979
  },
  // ... for targets 1-9
]

• Training parameters:
  • "is_finetuning": false - Initial training from scratch
  • "checkpoint_path": null - No pre-trained model

2. Fine-tuning

After multitask training, fine-tune the model for specific ADME parameters:

# Run fine-tuning with the same command
kmol train /path/to/ADME_MTFT/configs/accuracy_drug/adme/finetuning_learning_rate/train/clint/itr1/config.json

Key configuration features for fine-tuning:

• Single-target data: Uses target-specific CSV files (e.g., clint.csv for intrinsic clearance)

"loader": {
  "type": "multitask",
  "input_path": "clint.csv",
  "task_column_name": "target_id",
  "max_num_tasks": 10
}

• Pre-trained model loading:

"is_finetuning": true,
"checkpoint_path": "../ADMET/configs/accuracy_drug/adme/multitask/train/itr1/checkpoint.best.pt"

3. Prediction

To make predictions on new compounds:

# Run prediction using the fine-tuned model
kmol predict /path/to/ADME_MTFT/configs/accuracy_drug/adme/finetuning_learning_rate/test/clint/itr1/config.json

Key configuration features for prediction:

• Test data configuration:

"loader": {
  "input_path": "test/clint.csv"
},
"splitter": {
  "type": "index",
  "splits": {
    "test": 1.0
  }
}

• Fine-tuned model loading:

"is_finetuning": true,
"checkpoint_path": "../ADMET/configs/accuracy_drug/adme/finetuning_learning_rate/train/clint/itr1/checkpoint.best.pt"

• Same standardization parameters: Uses identical transformers as training to ensure consistent data preprocessing

• Evaluation metrics: Configured to compute R², MAE, and RMSE on test split

Notes

• Update the paths in the commands above to match your actual directory structure
• Ensure the kmol environment is activated before running any commands
• For detailed configuration options, refer to the kmol documentation