README

Project: Lightweight Deep Learning for Spatial Gene Expression Prediction

This repository contains the official implementation of our study on spatial gene expression prediction from single H&E histology spots using EfficientNet-B0 and lightweight deep learning strategies.

Our work establishes a reproducible performance ceiling for single-spot prediction tasks, significantly outperforming prior methods such as BLEEP and ST-Net on the GSE240429 human liver dataset【144†source】.

---

📌 Key Highlights

• Single-spot prediction pipeline (no contextual multi-spot input).

• EfficientNet-B0 backbone with full fine-tuning.

• Morphology-Preserving Augmentation (MPA): carefully designed augmentation strategy preserving biological structure.

• Benchmark results:

  • Up to 80% improvement over BLEEP on highly expressed genes (HEG).
  • Outperformed ResNet-50 while using only 1/5 of the parameters.

• Exploratory analyses:

  • Backbone comparison (ResNet, DenseNet, MobileNet, EfficientNet).
  • Augmentation comparison (ours vs. ST-Net).
  • Cluster-ID experiments.
  • Fine-tuning strategies (layer freezing, hidden layer depth).

---

📂 Repository Structure

BLEEP/
│── 19_make_patches_and_csv.py       # Patch extraction + dataset CSV generation
│── 20_train_effb0_sota_split8020.py # EfficientNet-B0 baseline training (80/20 split)
│── 21_train_backbones.py            # Backbone comparison (ResNet, DenseNet, etc.)
│── 22_search_freeze_and_hidden_cv5.py # 5-fold CV for layer freezing + hidden size search
│── 23_compare_augs.py               # Augmentation comparison (MPA vs ST-Net)
│── GSE240429_data/                  # Dataset folder (preprocessed data)
│── requirements.txt                 # Python dependencies
│── README.md                        # Project documentation
│── LICENSE                          # License file

---

🚀 Getting Started

1. Clone the repository

git clone https://github.com/Medical-AI-GSE240429/Code.git
cd your_repo

2. Install dependencies

pip install -r requirements.txt

3. Dataset

• Dataset: NCBI GEO: GSE240429
• Place the preprocessed dataset in GSE240429_data/ (TIFF files excluded).

4. Run patch extraction & dataset creation

python 19_make_patches_and_csv.py

5. Training & Evaluation

• EfficientNet-B0 baseline:

python 20_train_effb0_sota_split8020.py

• Backbone comparison:

python 21_train_backbones.py

• 5-fold CV + hyperparameter search:

python 22_search_freeze_and_hidden_cv5.py

• Augmentation comparison:

python 23_compare_augs.py

---

📊 Results (Summary)

• EfficientNet-B0 achieved r = 0.315 on HEG, significantly outperforming BLEEP and ST-Net【144†source】.
• Predicted 50 genes with r ≥ 0.30, compared to only 20 by ResNet-50 with 5× more parameters.
• MPA augmentation provided stable gains compared to ST-Net augmentations.

---

📜 Citation

If you use this repository, please cite:

@article{YourPaper2025,
  title   = {From Histology to Spatial Transcriptomics: Establishing a Lightweight Single-Patch Baseline},
  author  = {Hyungyum Jang†, Hyunsoo Shin†, Hawon Lee, Yena Jang, Sunghoon Jung*, Minji Jeon*},
  journal = {},
  year    = {2025}
}

---

🧑‍💻 Authors

• Hyungyum Jang† – [email protected]
• Hyunsoo Shin† – [email protected]
• Hawon Lee – [email protected]
• Yena Jang – [email protected]
• Sunghoon Jung* – [email protected]
• Minji Jeon* – [email protected]

Contributions follow the paper: †Equal contribution.

---

📄 License

This project is licensed under the terms of the MIT License.