InFo-DRP

InFo-DRP is a drug response prediction framework that improves robustness under chemical distribution shift by coupling invariant learning with foundation-model drug representations.

The model is evaluated on GDSCv1 and GDSCv2 benchmarks under drug-blind, cell-blind, scaffold-blind, and random-split settings.

This repository provides full training and evaluation code, predefined data splits, and trained model checkpoints used to report the results in the paper.

---

Repository Structure

InFo-DRP/
├── checkpoint/        # (Optional) Directory for storing downloaded trained model checkpoints
├── data/              # Processed GDSC data and predefined splits
│   ├── gdscv1/
│   └── gdscv2/
├── result/            # Evaluation outputs and prediction files
├── src/               # Source code
│   ├── dataset.py
│   ├── model.py
│   ├── util.py
│   └── main.py
├── environment.yml    # Conda environment specification
└── README.md

---

Environment Setup

We recommend using conda to reproduce the experimental environment.

conda env create -f environment.yml
conda activate infodrp

Verify PyTorch and CUDA:

python -c "import torch; print(torch.__version__, torch.cuda.is_available(), torch.version.cuda)"

---

GPU Selection

GPU selection must be performed before running the script by setting the CUDA_VISIBLE_DEVICES environment variable.

CUDA_VISIBLE_DEVICES=0 python src/main.py ...

---

Running Experiments

This section provides example commands for all generalization settings evaluated in the paper.

GDSCv1 Experiments

GDSCv1 is evaluated under drug-blind, cell-blind, and random-split settings.

• Drug-blind (newDRUG)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv1 \
  --split_mode newDRUG \
  --exp_name infodrp_gdscv1_newDRUG \
  --cuda

• Cell-blind (newCELL)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv1 \
  --split_mode newCELL \
  --exp_name infodrp_gdscv1_newCELL \
  --cuda

• Random-split (newPAIR)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv1 \
  --split_mode newPAIR \
  --exp_name infodrp_gdscv1_newPAIR \
  --cuda

---

GDSCv2 Experiments

GDSCv2 is evaluated under scaffold-blind, drug-blind, and cell-blind settings.

• Scaffold-blind (newSCAFFOLD)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv2 \
  --split_mode newSCAFFOLD \
  --exp_name infodrp_gdscv2_newSCAFFOLD \
  --cuda

• Drug-blind (newDRUG)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv2 \
  --split_mode newDRUG \
  --exp_name infodrp_gdscv2_newDRUG \
  --cuda

• Cell-blind (newCELL)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv2 \
  --split_mode newCELL \
  --exp_name infodrp_gdscv2_newCELL \
  --cuda

Additional training hyperparameters (e.g., learning rate, batch size, regularization coefficients) can be configured via command-line arguments. See python src/main.py --help for the full list of options.

---

Evaluation Only (Reproducing Reported Results)

CUDA_VISIBLE_DEVICES=0 python src/main.py \
  --dataset gdscv2 \
  --split_mode newSCAFFOLD \
  --exp_name repro_gdscv2_newSCAFFOLD \
  --eval_only \
  --ckpt_dir checkpoint/InFo-DRP_gdscv2_NewScaffold_best_ckpt \
  --cuda

---

Outputs

Results are saved under the result/ directory, including per-fold prediction CSV files and a summary JSON file reporting RMSE and PCC.

---

Checkpoints

Due to file size limitations, trained model checkpoints corresponding to the results reported in the paper are not included directly in this repository.

They can be downloaded from the following Google Drive link:

Google Drive – Trained Model Checkpoints

When running training commands, model checkpoints will be automatically saved under the checkpoint/ directory. The provided checkpoints can be optionally downloaded to directly evaluate the reported results without retraining.

---

Citation

Citation information will be added upon publication.