We introduce two metrics: EPS (Expression Predictability Score) and SPS (Slice Predictability Score), to quantify the predictability of gene expression from histology image. Python package expression_copilot is developed to calculate these metrics efficiently. It also provides several baseline models to predict gene expression from image embeddings, such as MLP and linear regression.

We recommend to install expression_copilot to a new conda environment:

conda create -n eps python=3.11 -y && conda activate eps
pip install expression_copilot

(Optional) If you have CUDA-enabled GPU, you could install cuml&cupy to accelerate KNN building, and install torch to accelerate MLP baseline training:

conda create -n eps_cuda -c conda-forge -c rapidsai -c nvidia python=3.11 rapids=25.06 'cuda-version>=12.0,<=12.8' -y && conda activate eps_cuda
pip install expression_copilot[torch]

You could also use our pre-built docker image directly:

# GPU version
docker run --gpus all -it --rm huhansan666666/expression_copilot:latest

# CPU version
docker run -it --rm huhansan666666/expression_copilot:latest

The following code snippet shows how to calculate EPS and SPS via expression_copilot package. We assume you have already preprocessed your spatial transcriptomics data into an AnnData object (adata), where adata.X should store raw counts and adata.obsm['IMAGE_KEY_NAME'] should store image embeddings of spots. (Preprocessed steps are described in Advanced Tutorial in detail)

import scanpy as sc
import numpy as np
from expression_copilot import ExpressionCopilotModel

# Load data
# adata.X is raw counts
# adata.obsm['X_uni'] stores image embeddings of spots
url = 'https://drive.google.com/uc?id=10WD9vFgsoMoTt6g3017XxNK_bq8qp3oM'
adata = sc.read('./adata_with_image_emb.h5ad', backup_url=url)

# Init model
model = ExpressionCopilotModel(adata, image_key = 'X_uni')

# Calculate EPS and SPS
eps = model.calc_metrics_per_gene()
sps = eps.mean()

# Run baseline model (support 'ridge', 'linear', 'ensemble', 'mlp')
baseline_metrics_per_gene, _ = model.calc_baseline_metrics(method = 'mlp')

We provide several tutorials in the resource/tutorials folder. You could also run them in Google Colab directly:

Name	Description	Colab
Basic Tutorial	Basic tutorial of calculating EPS
Advanced Tutorial	Start with 10x spatial-ranger output from scratch
Multi-omics Tutorial	Calculating EPS and SPS on single cell multi-omics data

In coming.

If you want to repeat results in the manuscript, please check the experiments folder.

Please open a new github issue if you have any question.

1. numba related bugs

We use numba to increase the speed (up to 12x). However, it may have compatibility issues with different python/numpy versions. We tested the latest version of numba (0.6.12) and it works fine with Python 3.11/3.12, numpy 1.26.

We thank the following great open-source projects for their help or inspiration:

• HEST
• UNI
• entropy_estimators