SCG: Spatially Co-expressed Gene Identification

High-resolution, spatially varying gene co-expression networks for spatial transcriptomics

Overview

SCG identifies gene pairs whose co-expression relationship changes meaningfully across a tissue section — what we call Spatially Co-expressed Genes (SCGs). Rather than estimating a single network shared across all locations, SCG estimates spot-specific co-expression networks that vary continuously across the tissue, without requiring pre-defined regions or clusters.

SCG is built on the Spatial Covariance Regression (SCR) framework: a spatially varying Bayesian factor model with GPU-accelerated coordinate-ascent variational inference that scales to modern platforms such as 10x Genomics Xenium (>50,000 spots).

**(A)** SCG decomposes gene expression into spatial bases and low-dimensional factors, yielding a parsimonious spatially varying covariance model. **(B)** SCGs are identified via Otsu thresholding on posterior spatial standard deviations, with FDR control at $\alpha$ = 0.1. **(C)** SCG outperforms spCorr on MCC and FPR across all simulated sample sizes (5K–50K spots); spCorr fails to scale beyond 20K. **(D–F)** Application to human breast cancer (Visium) recovers pathway-level network rewiring across tissue regions, with spatially varying correlation fields highlighting co-expression patterns in angiogenesis and tumor-promoting inflammation pathways.

Key Features

Estimates spot-level co-expression networks
Scales to Visium, Xenium, and other high-resolution ST platforms
Rigorous uncertainty quantification via variational posterior
Data-driven SCG identification with FDR control at 10%
Hardware-agnostic: runs on CPU or GPU

Performance

SCG demonstrates strong and consistent recovery of spatially varying network structure across all spatial resolutions tested (5K–50K spots). Mean MCC ranged from 0.92 to 0.94, with TPR between 0.96 and 0.99 and FPR below 0.02 throughout, indicating reliable discrimination despite the class imbalance inherent to sparse networks. Against spCorr, SCG exceeded MCC by 0.10 to 0.20 at every matched sample size, ran five to seven times faster per edge, and scaled to resolutions where spCorr failed entirely.

Metric	SCG	spCorr
MCC	0.92–0.94	0.72–0.83
TPR	0.97–0.98	0.81–1.00
FPR	0.01–0.02	0.04–0.06
Max spots	50K+	~20K

Applications

SCG has been applied to:

Breast cancer (10x Visium, ~4,700 spots): recovers pathway-level network rewiring across tumor, intermediate, and normal tissue regions
Alzheimer’s disease (10x Xenium, ~57,000 spots): replicates and extends the plaque-induced gene (PIG) co-expression module across anatomical regions

Installation

Requires Python $\geq$ 3.10. GPU support via CUDA is optional but recommended for datasets >20K spots.

pip install git+https://github.com/iebuker/SCG.git

Citation

Buker, I. E., Ni, Y., Hicks, S. C., Kang, J., & Acharyya, S. (2026). SCG: Spatially co-expressed gene identification through spatially varying networks [Preprint]. bioRxiv.

--- title: "SCG: Spatially Co-expressed Gene Identification" subtitle: "High-resolution, spatially varying gene co-expression networks for spatial transcriptomics" --- ## Overview **SCG** identifies gene pairs whose co-expression relationship changes meaningfully across a tissue section — what we call **Spatially Co-expressed Genes (SCGs)**. Rather than estimating a single network shared across all locations, SCG estimates spot-specific co-expression networks that vary continuously across the tissue, without requiring pre-defined regions or clusters. SCG is built on the **Spatial Covariance Regression (SCR)** framework: a spatially varying Bayesian factor model with GPU-accelerated coordinate-ascent variational inference that scales to modern platforms such as 10x Genomics Xenium (>50,000 spots). ![**(A)** SCG decomposes gene expression into spatial bases and low-dimensional factors, yielding a parsimonious spatially varying covariance model. **(B)** SCGs are identified via Otsu thresholding on posterior spatial standard deviations, with FDR control at $\alpha$ = 0.1. **(C)** SCG outperforms spCorr on MCC and FPR across all simulated sample sizes (5K–50K spots); spCorr fails to scale beyond 20K. **(D–F)** Application to human breast cancer (Visium) recovers pathway-level network rewiring across tissue regions, with spatially varying correlation fields highlighting co-expression patterns in angiogenesis and tumor-promoting inflammation pathways.](images/combined_figure_1.png) ## Key Features - Estimates spot-level co-expression networks - Scales to Visium, Xenium, and other high-resolution ST platforms - Rigorous uncertainty quantification via variational posterior - Data-driven SCG identification with FDR control at 10% - Hardware-agnostic: runs on CPU or GPU ## Performance SCG demonstrates strong and consistent recovery of spatially varying network structure across all spatial resolutions tested (5K–50K spots). Mean MCC ranged from 0.92 to 0.94, with TPR between 0.96 and 0.99 and FPR below 0.02 throughout, indicating reliable discrimination despite the class imbalance inherent to sparse networks. Against [spCorr](https://github.com/chexjiang/spCorr), SCG exceeded MCC by 0.10 to 0.20 at every matched sample size, ran five to seven times faster per edge, and scaled to resolutions where spCorr failed entirely. | Metric | SCG | spCorr | |--------|-----|--------| | MCC | 0.92–0.94 | 0.72–0.83 | | TPR | 0.97–0.98 | 0.81–1.00 | | FPR | 0.01–0.02 | 0.04–0.06 | | Max spots | 50K+ | ~20K | ## Applications SCG has been applied to: - **Breast cancer** (10x Visium, ~4,700 spots): recovers pathway-level network rewiring across tumor, intermediate, and normal tissue regions - **Alzheimer's disease** (10x Xenium, ~57,000 spots): replicates and extends the plaque-induced gene (PIG) co-expression module across anatomical regions ## Installation Requires Python $\geq$ 3.10. GPU support via CUDA is optional but recommended for datasets >20K spots. ```bash pip install git+https://github.com/iebuker/SCG.git ``` ## Citation Buker, I. E., Ni, Y., Hicks, S. C., Kang, J., & Acharyya, S. (2026). [*SCG: Spatially co-expressed gene identification through spatially varying networks*](https://doi.org/10.64898/2026.09.01.748618) [Preprint]. bioRxiv.