UniCure is the first multimodal foundation model integrating omics (UCE) and chemical (Uni-mol) foundation models to predict transcriptomic drug responses across diverse cellular contexts. Trained on 1.8M+ perturbation profiles (22k compounds, 166 cell types, 24 tissues).

💡 Key Innovations

• FlexPert: Sliding-window cross-attention for flexible drug-cell interaction modeling
• LoRA-PEFT: Parameter-efficient tuning preserving pretrained knowledge
• MMD loss: Handles unpaired perturbation data without cell matching
• Staged training: Enhancing computational efficiency and functional modularity

OS: Linux (Ubuntu 22.04 LTS or Rocky Linux 8.6+ recommended)
Environment Manager: Miniconda/Mamba

# Base environment (minimal)
conda env create -f environment.yml

# Full environment (with development tools)
conda env create -f environment_full.yml

1. Install Python 3.10
conda create -n unicure python=3.10
conda activate unicure

2. Install PyTorch (select appropriate CUDA version)
⚠ Check latest at: https://pytorch.org/get-started/locally/
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

3. Install Accelerate & DeepSpeed (recommended for reproduction)
pip install accelerate
⚠ Follow configuration (about how to configure DeepSpeed): https://github.com/huggingface/accelerate

4. Install core dependencies
pip install numpy pandas scikit-learn fastparquet tqdm anndata scanpy lora-pytorch

5. Install Uni-Mol (required for testing)
You can create a new conda environment. 
https://github.com/deepmodeling/Uni-Mol

Download and overwrite these folders to your local UniCure directories:

1. data folder

   • Contains: LINCS, SciPlex datasets, and PTC
   • Local path: your_project_path/UniCure/data/

2. requirement folder

   • Contains: Pre-trained model weights and configuration files
   • Local path: your_project_path/UniCure/requirement/

⚠️ Overwrite Notice: Replace existing directories completely when copying
⚠️ Unzip Notice: Unzip requirement/model_weights/best_model.rar

Download these essential files to requirement/UCE_pretraining_files/:

After downloading, confirm directory structure:

UniCure/
├── data/
│   ├── lincs2020/
│   ├── sciplex/ 
│	└── PTC/
└── requirement/
    ├── model_weights/
    └── UCE_pretraining_files/
        ├── 33l_8ep_1024t_1280.torch
        ├── all_tokens.torch
        └── species_chrom.csv

python start_test.py

python main.py

Zexi Chen 📧 Email: [email protected]

doi: https://doi.org/10.1101/2025.06.14.658531

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.