Get Started

GPU-accelerated spatial omics analysis framework

Overview

spatial-gpu is a GPU-accelerated spatial omics analysis framework that serves as a drop-in replacement for Squidpy, delivering 10–100x speedup on common spatial transcriptomics workflows.

The package provides a comprehensive Python reimplementation of the SpaCET cell type deconvolution algorithm and the SecAct secreted protein activity inference pipeline, plus spatial graph construction, neighborhood enrichment, co-occurrence analysis, and cell segmentation integration.

Key features:

Two-stage hierarchical cell type deconvolution (SpaCET algorithm)
Secreted protein activity inference via ridge regression (SecAct)
Signaling pattern discovery, velocity visualization, and spatial CCC
Gene set scoring (Hallmark, CancerCellState, TLS)
Spatial autocorrelation (Moran’s I) and ligand-receptor analysis
Multi-backend CPU/GPU switching (NumPy / CuPy / RAPIDS)
Memory-efficient chunked processing for 100M+ cell datasets
Readers for Visium, Xenium, CosMx, and MERSCOPE platforms

Installation

Install the latest release from PyPI:

pip install spatialgpu

For development (editable install with dev dependencies):

git clone https://github.com/psychemistz/spatial-gpu.git
cd spatial-gpu
pip install -e ".[dev]"

GPU acceleration requires a CUDA toolkit. Install the appropriate CuPy backend for your CUDA version:

# CUDA 12
pip install cupy-cuda12x

# CUDA 11
pip install cupy-cuda11x

Quick Start

Below is a minimal example that loads 10X Visium data, runs cell type deconvolution, and visualizes the results.

import spatialgpu.deconvolution as spacet

# 1. Load 10X Visium data from Space Ranger output
adata = spacet.create_spacet_object_10x("path/to/spaceranger/outs")

# 2. Run two-stage hierarchical deconvolution
adata = spacet.deconvolution(adata, cancer_type="BRCA")

# 3. Visualize cell type fractions
spacet.visualize_spatial_feature(
    adata,
    spatial_type="CellFraction",
    spatial_features=["Malignant", "Macrophage"],
)

# 4. Infer secreted protein activity
adata = spacet.secact_inference(adata)

# 5. Visualize signaling velocity
spacet.visualize_secact_velocity(adata, gene="SPP1")

Tutorials

Explore the full set of tutorials for detailed, end-to-end workflows:

Visium Breast Cancer — Full deconvolution + cell-cell interaction workflow
Old ST PDAC — Deconvolution with matched scRNA-seq reference
Hi-Res ST CRC — High-resolution spatial transcriptomics deconvolution
Gene Set Score — Gene set scoring (Hallmark, CancerCellState, TLS)
Spatial Correlation — Moran’s I spatial autocorrelation analysis
Signaling Patterns — SecAct activity + signaling patterns + velocity
Cell-Cell Communication — Single-cell resolution spatial CCC
Bulk Benchmark — Pseudobulk deconvolution evaluation + MuSiC/CIBERSORTx export

For the full API listing, see the Reference page.