STaRN

Summary Table Representation Learning with Semantic Neighborhood Distillation

STaRN is a self-supervised framework for learning radiomics representations of spatial transcriptomics (ST) spots. The core idea is to treat each training batch as a Summary Table — rows are spots (anchor + spatial KNN neighbours + random globals), columns are radiomics features — and apply a SAINT-style dual-attention encoder that models both within-spot feature interactions (column attention) and cross-spot neighbourhood context (row attention). Representations are distilled toward a frozen UNI pathology foundation model, which provides semantic ordering signals without being used as a feature extractor itself.

---

Overview

---

Installation

conda create -n starn python=3.10
conda activate starn

Install PyTorch matching your CUDA version:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Install remaining dependencies:

pip install timm h5py scanpy scikit-learn

---

Run

1. Extract UNI embeddings (once, before training)

python .test/extract_uni.py

2. Edit config — set data_root, sample_ids, and hyperparameters in configs/config.py.

3. Train

python train.py