Visium Human Breast Cancer analysis using Spatial VAE

Quick Start

mkdir data
cd data

Then, download the Visium_Human_Breast_Cancer Dataset using the following commands

curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_molecule_info.h5
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_filtered_feature_bc_matrix.h5
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_filtered_feature_bc_matrix.tar.gz
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_raw_feature_bc_matrix.h5
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_raw_feature_bc_matrix.tar.gz
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_analysis.tar.gz
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_spatial.tar.gz
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_spatial_enrichment.csv
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_metrics_summary.csv
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_web_summary.html
curl -O https://cf.10xgenomics.com/samples/spatial-exp/1.3.0/Visium_Human_Breast_Cancer/Visium_Human_Breast_Cancer_cloupe.cloupe

Then, to install packages, run

uv sync

Finally, run the following to train the model

uv run main.py

Loss function

We utlize the following loss function in this work for the spatial VAE:

$$ \mathcal{L}

|x - \hat{x}|2^2 ;+; \beta \cdot \tfrac{1}{2} \left(\mu^2 + e^{\log\sigma^2} - 1 - \log\sigma^2\right) ;+; \lambda{\text{spatial}} \cdot \frac{1}{|E|}\sum_{(u,v)\in E} | z_u - z_v |_2^2 $$

Pipeline

Expression is log normalized, and spatial features are engineered from coordinates by computing the fraction of nearest k neighbors of a given cell type.

Results

We observe distinct clusters and spatial element classification as indicated in the spatial clusters

We also observe proper latent space clustering