MMAP: A Multi-Magnification and Prototype-Aware Architecture for Predicting Spatial Gene Expression

Link to paper: https://arxiv.org/abs/2510.11344

Here, we reimplemented a number of different baselines models for a task related to Spatial Transcriptomics (ST) - predicting gene expression from Whole Slide Images (WSIs) Original source code for our proposed method will be updated later

Models implemented so far:

• ST-Net
• HisToGene
• DeepPT
• TCGN

Installation guide

• Download data: Run "git clone https://github.com/almaan/her2st.git" in the "data" folder to clone the HER2ST dataset
• To unzip files in ST-cnts, cd into "MMAP-Baselines/data/her2st/data/ST-cnts/" and run "gunzip *.gz"
• Run "pip install -r requirements.txt" to download all the necessary Python libraries (with the versions we are using)

Note: You may encounter installation issues. In such case, refer to the Guide for Installation or Compatibility issues (.txt file)

Usage

For ST-Net, HisToGene, TCGN:

1. Train: Run ST_train.py and choose the model you wish to use. The model will conduct training and save checkpoints once completed.
2. Test: Run ST_predict.py and choose the model you wish to use. Results printed are shown in metrics (printed in cmd line) and figures (imgs saved in the "figures" folder)

For DeepPT: (To be updated)

References and links to source codes

ST-Net

Integrating spatial gene expression and breast tumour morphology via deep learning
by Bryan He, Ludvig Bergenstråhle, Linnea Stenbeck, Abubakar Abid, Alma Andersson, Åke Borg, Jonas Maaskola, Joakim Lundeberg & James Zou.
Nature Biomedical Engineering (2020).

Since the code provided in their GitHub has been deprecated, we referred to the re-implemented versions from authors of HisToGene to reproduce our own re-implemented version of ST-Net

HisToGene

Leveraging information in spatial transcriptomics to predict super-resolution gene expression from histology images in tumors
by Minxing Pang, Kenong Su, Mingyao Li.
biorxiv (2021).

DeepPT

A deep-learning framework to predict cancer treatment response from histopathology images through imputed transcriptomics
by Hoang, DT., Dinstag, G., Shulman, E.D. et al.
Nature Cancer (2024).

TCGN

Transformer with Convolution and Graph-Node co-embedding: A accurate and interpretable vision backbone for predicting gene expressions from local histopathological image
by Xiao Xiao, Yan Kong, Zuoheng Wang, Hui Lu.
Medical Image Analysis (2024).