---

ClinDRP: Multimodal Drug Response Prediction

Accurate identification of drug response in cancer patients is the foundation of precision medicine. Recently, molecular language models provide the great opportunities for drug response prediction. However, most methods ignore the mapping relation cross modalities, leading to deviate from the basic criterion of molecular modeling. Furthermore, the clinical response of patients is closely associated with multi-scale regulatory networks within life systems. In this study, we propose a deep learning framework, ClinDRP, that integrates cross-modal interaction of molecules and multi-order regulatory within patient to improve clinical drug response prediction. Specifically, ClinDRP designs a conditional masked language model on the unified multimodal sequences where three-dimensional (3D) conformations are discretized into tokens as context of 1D sequence, thus capturing fine-grained interaction cross modalities. In patient representation learning, regulatory networks at gene, protein and pathway levels are hierarchically embedded into Transformer through adaptive mapping functions. Across multiple scenarios of drug response and disease progression prediction, ClinDRP achieved superior performance and explored potential mechanisms behind pharmacological and biochemical processes. Importantly, the sensitive drugs predicted by ClinDRP can improve the survival outcomes of clinical patients, further highlighting its potential in anticancer drug therapies.

🛠 Installation

Install the required dependencies using requirements.txt:

pip install -r requirements.txt

---

📂 Project Structure

To ensure the code runs seamlessly with default parameters, it is recommended to maintain the following directory structure:

Project/
├── main.py                # Entry point containing launch logic for all tasks
├── model/                 # Stores pretrained model weights and tokenizer
│   ├── model_path/
│   │   └── checkpoint-80000/
│   └── tokenizer/
├── data/                  # Data directory
│   ├── PDTC/              # PDTC related data (.csv, .npy)
│   └── PDX/               # PDX related data (.csv, .npy)
└── README.md

---

🚀 Usage

This project uses a unified main.py entry point and utilizes Subcommands to distinguish between different experimental tasks.

The code is designed to support two execution modes:

1. Reproduction Mode: Run without arguments to automatically load the preset default paths and hyperparameters defined in the code.
2. Custom Experiment Mode: Override default configurations via command-line arguments.

1. Pretraining

Use this task if you wish to re-pretrain the model.

• Default Execution:

  python main.py Pretrain

• Specify Pretraining Data:

  python main.py Pretrain \
      --train_data 'data/Pretrain/out_embedding_train.csv'

2. TransferEvaluation Task

Runs the baseline PDTC dataset training task. By default, it trains using Sample data and predicts outcomes on Model data.

• Default Execution:

  python main.py pdtc-reg

• Custom Execution:

  # Set batch size to 64, learning rate to 5e-5, and repeat 3 times
  python main.py pdtc-reg \
      --batch_size 64 \
      --learning_rate 5e-5 \
      --repeat_times 3

3. PDTC CrossValidation

Runs a rigorous 10-fold cross-validation experiment to evaluate the model's generalization capability.

• Default Execution:

  python main.py pdtc-ten

• Custom Execution:

  # Specify WandB project name and increase number of Epochs
  python main.py pdtc-ten \
      --project_name "PDTC_10Fold_Exp1" \
      --num_epochs 200 \
      --batch_size 128

4. PDX CrossValidation

Runs a rigorous 10-fold cross-validation experiment to evaluate the model's generalization capability.

• Default Execution:

  python main.py pdx-clas

• Custom Execution:

  # Specify WandB project name and increase number of Epochs
  python main.py pdx-clas \
      --project_name "PDX_10Fold_Exp1" \
      --num_epochs 200 \
      --batch_size 128

---

⚙️ Arguments

You can view the full list of supported parameters for any task by running python main.py <task> --help. Below are the common arguments:

Argument	Type	Description	Default Value
--model_path	str	Path to the pretrained model	./model/multiSmiles_model/...
--tokenizer_path	str	Path to the tokenizer	./model/custom_tokenizer
--batch_size	int	Training batch size	128
--learning_rate	float	Initial learning rate	1e-4 or 3e-5
--num_epochs	int	Number of training epochs	150
--dropout_rate	float	Dropout probability
--project_name	str	WandB project name	(Named automatically based on task)
--weight_decay	float	Weight decay coefficient	0.0001 / 0.001

---

📄 Citation

If you use this project in your research, please cite the following reference: