GeneFlow

Official implementation of GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025).

---

News

• November 2025: Initial release of GeneFlow codebase
• October 2025: Preprocessed demo Xenium samples released on Zenodo
• September 2025: Paper accepted at NeurIPS 2025

---

Overview

GeneFlow translates single-cell gene expression profiles to histopathological images using rectified flow models. This approach enables the generation of synthetic tissue images from spatial transcriptomics data, facilitating downstream analyses in computational pathology and spatial biology.

---

Installation

Requirements

• CUDA 12.1 compatible GPU (tested on 8×H100 with DDP and 1×H100)
• Conda or Miniconda
• Python 3.11

Setup

1. Create conda environment:

conda create -n geneflow python=3.11
conda activate geneflow

2. Install PyTorch with CUDA 12.1:

conda install pytorch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 pytorch-cuda=12.1 -c pytorch -c nvidia

3. Install dependencies:

pip install -r requirements.txt

4. Verify installation:

python -c "import torch; print(f'PyTorch: {torch.__version__}'); print(f'CUDA available: {torch.cuda.is_available()}')"

---

Data

Download Preprocessed Demo Data

We provide three preprocessed Xenium samples for demonstration, which can be downloaded from https://zenodo.org/records/17429142:

Sample C1:

wget -O xenium_c1.tar.gz "https://zenodo.org/records/17429142/files/Xenium_V1_hSkin_Melanoma_Base_FFPE.tar.gz?download=1"
tar -xzf xenium_c1.tar.gz -C /path/to/GeneFlow/processed_data/

Sample C2:

wget -O xenium_c2.tar.gz "https://zenodo.org/records/17429142/files/Xeniumranger_V1_hSkin_Melanoma_Add_on_FFPE.tar.gz?download=1"
tar -xzf xenium_c2.tar.gz -C /path/to/GeneFlow/processed_data/

Sample P1:

wget -O xenium_p1.tar.gz "https://zenodo.org/records/17429142/files/Xenium_Prime_Human_Skin_FFPE.tar.gz?download=1"
tar -xzf xenium_p1.tar.gz -C /path/to/GeneFlow/processed_data/

Data Preprocessing

For preprocessing custom Xenium data from HEST-1k:

python utils/prepare_hestxenium_data.py \
    --input_dir /path/to/GeneFlow/raw/data \
    --output_dir /path/to/GeneFlow/processed_data/ \
    --img_size 256 \
    --img_channels 4

---

Training

Run training with default parameters:

bash train.sh

Custom Training

Modify train.sh or run directly with custom arguments:

python rectified/rectified_main.py \
    --model_type single \
    --adata /path/to/GeneFlow/processed_data/adata.h5ad \
    --image_paths /path/to/GeneFlow/processed_data/cell_image_paths.json \
    --img_size 256 \
    --img_channels 4 \
    --output_dir /path/to/GeneFlow/results \
    --batch_size 16 \
    --epochs 50 \
    --patience 5 \
    --lr 1e-4 \
    --use_amp

Distributed Training

For multi-GPU training:

bash train_distributed.sh

Or run directly:

torchrun --nproc_per_node=8 rectified/rectified_main.py \
    --use_ddp \
    --use_amp \
    --model_type single \
    --adata /path/to/GeneFlow/processed_data/adata.h5ad \
    --image_paths /path/to/GeneFlow/processed_data/cell_image_paths.json \
    --img_size 256 \
    --img_channels 4 \
    --output_dir /path/to/GeneFlow/results \
    --batch_size 16 \
    --epochs 50

---

Generation

Generate synthetic histopathological images from gene expression using a pretrained model:

bash generate.sh

Custom Generation

Modify generate.sh or run directly with custom arguments:

python rectified/rectified_generate.py \
    --model_path /path/to/GeneFlow/results/checkpoints/best_model.pt \
    --model_type single \
    --adata /path/to/GeneFlow/processed_data/adata.h5ad \
    --image_paths /path/to/GeneFlow/processed_data/cell_image_paths.json \
    --img_size 256 \
    --img_channels 4 \
    --output_dir /path/to/GeneFlow/generated_results \
    --batch_size 8 \
    --num_samples 100 \
    --gen_steps 50

Generation Parameters

• --model_path: Path to pretrained model checkpoint
• --num_samples: Number of samples to generate
• --gen_steps: Number of generation steps for rectified flow (default: 50)
• --enable_stain_normalization: Apply stain normalization to generated images
• --stain_normalization_method: Method for stain normalization (default: skimage_hist_match)

Output

Generation produces:

• Individual PNG images for each sample (RGB and auxiliary channels)
• Multi-page PDF with all samples (generation_results.pdf)
• Preview PNG with first 10 samples (generation_results.png)

---

Evaluation

Run comprehensive evaluation on trained model:

bash eval.sh

Custom Evaluation

Modify eval.sh or run directly with custom arguments:

python rectified/rectified_evaluate.py \
    --model_path /path/to/GeneFlow/results/checkpoints/best_model.pt \
    --model_type single \
    --adata /path/to/GeneFlow/processed_data/adata.h5ad \
    --image_paths /path/to/GeneFlow/processed_data/cell_image_paths.json \
    --img_size 256 \
    --img_channels 4 \
    --output_dir /path/to/GeneFlow/evaluation_results \
    --batch_size 8 \
    --gen_steps 50 \
    --use_amp

Distributed Evaluation

For multi-GPU evaluation:

bash eval_distributed.sh

Or run directly:

torchrun --nproc_per_node=8 rectified/rectified_evaluate.py \
    --use_ddp \
    --use_amp \
    --model_path /path/to/GeneFlow/results/checkpoints/best_model.pt \
    --model_type single \
    --adata /path/to/GeneFlow/processed_data/adata.h5ad \
    --image_paths /path/to/GeneFlow/processed_data/cell_image_paths.json \
    --img_size 256 \
    --img_channels 4 \
    --output_dir /path/to/GeneFlow/evaluation_results \
    --batch_size 8 \
    --gen_steps 50

Evaluation Metrics

GeneFlow evaluation includes:

Image Quality Metrics:

• SSIM (Structural Similarity Index)
• PSNR (Peak Signal-to-Noise Ratio)
• FID (Fréchet Inception Distance)

Biological Validation Metrics:

• UNI2-h FID (using histopathology foundation model)
• Cell type classification accuracy
• Nuclear feature similarity
• Spatial pattern similarity
• UNI2-h embedding similarity

RNA Prediction Metrics (Round-trip validation):

• Gene-wise correlation
• Sample-wise correlation
• RNA MSE

Biological Feature Evaluation

For comprehensive evaluation using biological features, install the required models:

Required Tools

1. Sequoia: Cell segmentation and feature extraction

   • Repository: https://github.com/gevaertlab/sequoia-pub
   • Installation: pip install sequoia-pub

2. UNI2: Universal histopathology foundation model

   • Model: https://huggingface.co/MahmoodLab/UNI2-h

3. HE2RNA Weights: Pretrained histology-to-transcriptomics model

   • Weights: https://huggingface.co/gevaertlab

Setup Instructions

1. Download HE2RNA weights:

wget -O sequoia/models/he2rna_weights.pt "https://huggingface.co/gevaertlab/he2rna/resolve/main/weights.pt"

2. Modify Sequoia imports in sequoia/src/he2rna.py:

# Replace:
# from src.read_data import SuperTileRNADataset
# from src.utils import patient_split, patient_kfold, custom_collate_fn, filter_no_features

# With:
from .read_data import SuperTileRNADataset
from .utils import patient_split, patient_kfold, custom_collate_fn, filter_no_features

3. Run evaluation with biological models:

python rectified/rectified_evaluate.py \
    --model_path /path/to/GeneFlow/results/checkpoints/best_model.pt \
    --model_type single \
    --adata /path/to/GeneFlow/processed_data/adata.h5ad \
    --image_paths /path/to/GeneFlow/processed_data/cell_image_paths.json \
    --output_dir /path/to/GeneFlow/evaluation_results \
    --uni2h_model_path /path/to/uni2h_model \
    --he2rna_weights /path/to/he2rna_weights.pt \
    --save_embeddings \
    --embeddings_output_path /path/to/embeddings

Output

Evaluation produces:

• evaluation_summary.json: Overall metrics summary
• per_sample_metrics.csv: Per-sample detailed metrics
• batch_level_biological_validation.csv: Batch-level biological metrics
• Visualization plots for all metrics
• UNI2-h embeddings (if --save_embeddings is enabled)

---

Citation

If you find this work useful, please cite:

@inproceedings{wanggeneflow,
  title={GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow},
  author={Wang, Mengbo and Verma, Shourya and Malusare, Aditya and Wang, Luopin and Lu, Yiyang and Aggarwal, Vaneet and Sola, Mario and Grama, Ananth and Lanman, Nadia Atallah},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025}
}

---

License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License.

Permitted:

• Use for academic and research purposes
• Citation in academic publications

Prohibited:

• Commercial use
• Distribution of modified versions
• Use in production systems without explicit permission

---

Contact

For questions, issues, or collaboration inquiries:

• Open an issue on this repository
• Check existing issues before creating new ones
• Provide detailed information for bug reports (OS, CUDA version, error messages)

---

Acknowledgments

We thank the developers of HEST-1k, Xenium platform, and the computational pathology community for their foundational contributions to spatial biology.