SCORPION

Documentation: https://zaoqu-liu.github.io/SCORPION/

What's New 🎉

v1.2.2 (January 2026)

🚀 RcppArmadillo acceleration: C++ implementation of core algorithms with BLAS optimization (~2.5x faster)
📚 Documentation website: Full pkgdown site with tutorials at https://zaoqu-liu.github.io/SCORPION/
📖 Vignettes: 4 comprehensive guides (Quick Start, Algorithm, Advanced Usage, Visualization)
⚡ Performance: Vectorized operations + optimized BLAS for matrix computation
🛡️ Robustness: Enhanced numerical stability in Tanimoto similarity
🌐 R-universe: Easy installation via install.packages("SCORPION", repos = "https://zaoqu-liu.r-universe.dev")

Overview

SCORPION (Single-Cell Oriented Reconstruction of PANDA Individually Optimized Gene Regulatory Networks) is a computational framework for inferring cell-type-specific gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data.

Motivation

Single-cell transcriptomics has revolutionized our understanding of cellular heterogeneity, yet the inherent sparsity and technical noise in scRNA-seq data pose significant challenges for accurate gene regulatory network inference. SCORPION addresses these limitations through a metacell-based coarse-graining strategy that preserves biological signal while reducing technical artifacts.

Method

SCORPION integrates three complementary data modalities to reconstruct transcriptome-wide GRNs:

Transcription factor (TF) binding motifs - sequence-based prior knowledge of potential regulatory relationships
Protein-protein interactions (PPI) - physical interactions between transcription factors
Gene expression profiles - cell-type-specific transcriptional states

The algorithm employs the PANDA (Passing Attributes between Networks for Data Assimilation) message-passing framework, which iteratively refines network edge weights through Tanimoto similarity-based updates until convergence.

SCORPION method overview showing the PANDA message-passing algorithm for gene regulatory network inference

Key Features

Metacell aggregation: Reduces data sparsity by grouping transcriptionally similar cells using graph-based clustering in PCA space
Multi-modal integration: Combines transcriptomic data with prior biological knowledge (TF motifs, PPI networks)
Comparable networks: Generates fully-connected, weighted, and directed GRNs suitable for population-level comparative analyses
Scalable: Efficient implementation supporting both CPU and GPU computation

Installation

From R-universe (Recommended)

install.packages("SCORPION", repos = "https://zaoqu-liu.r-universe.dev")

From CRAN

install.packages("SCORPION")

From GitHub (Development Version)

if (!requireNamespace("remotes", quietly = TRUE))
    install.packages("remotes")
remotes::install_github("Zaoqu-Liu/SCORPION")

Quick Start

library(SCORPION)

# Load example data
data(scorpionTest)

# Infer gene regulatory network
grn <- scorpion(
    tfMotifs = scorpionTest$tf,      # TF-target motif prior
    gexMatrix = scorpionTest$gex,    # Gene expression matrix (genes × cells)
    ppiNet = scorpionTest$ppi,       # Protein-protein interactions
    alphaValue = 0.1                 # Learning rate (default)
)

Input Data Format

Parameter	Description	Format
`tfMotifs`	TF-target regulatory prior	data.frame with columns: TF, Target, Score
`gexMatrix`	Gene expression matrix	matrix/dgCMatrix (genes × cells)
`ppiNet`	Protein-protein interactions	data.frame with columns: TF1, TF2, Score

Output

SCORPION returns a list containing:

Component	Description	Dimensions
`regNet`	TF-gene regulatory network	TFs × Genes
`coregNet`	Gene co-regulatory network	Genes × Genes
`coopNet`	TF cooperative network	TFs × TFs

Edge weights represent Z-scores indicating the strength of regulatory relationships.

Example Workflow

# Load package and data
library(SCORPION)
data(scorpionTest)

# Run SCORPION
result <- scorpion(
    tfMotifs = scorpionTest$tf,
    gexMatrix = scorpionTest$gex,
    ppiNet = scorpionTest$ppi,
    gammaValue = 10,        # Metacell graining level
    nPC = 25,               # Principal components for clustering
    alphaValue = 0.1,       # Learning rate
    hammingValue = 0.001    # Convergence threshold
)

# Examine output
dim(result$regNet)    # Regulatory network dimensions
# [1] 783 214

# Extract top regulatory edges
regNet <- result$regNet
top_edges <- which(abs(regNet) > 2, arr.ind = TRUE)

Parameters

Parameter	Default	Description
`gammaValue`	10	Ratio of cells to metacells
`nPC`	25	Number of principal components for kNN graph
`alphaValue`	0.1	Learning rate for network updates
`hammingValue`	0.001	Convergence threshold
`assocMethod`	"pearson"	Association method ("pearson", "spearman", "pcNet")

Performance

SCORPION v1.2.2 includes significant performance optimizations through RcppArmadillo and BLAS acceleration.

Benchmark Results

Metric	Before Optimization	After Optimization	Speedup
Test data (230 genes × 80 cells)	~10 sec	~4 sec	2.5x
1000×1000 matrix multiplication	~230 ms	11 ms	21x
Tanimoto similarity computation	~230 ms	19 ms	12x

Optimization Features

Feature	Description
RcppArmadillo	Core Tanimoto algorithm implemented in C++
BLAS Acceleration	Leverages system-optimized BLAS (OpenBLAS, Apple Accelerate, MKL)
Vectorization	All R operations use vectorized implementations
Sparse Matrix	Native support for sparse matrices via Matrix package

Enable Apple Accelerate (macOS)

For optimal performance on macOS, enable Apple's vecLib BLAS:

cd /Library/Frameworks/R.framework/Versions/4.4-arm64/Resources/lib
sudo ln -sf libRblas.vecLib.dylib libRblas.dylib

Enable OpenBLAS (Linux)

# Ubuntu/Debian
sudo apt-get install libopenblas-dev
sudo update-alternatives --config libblas.so.3

Estimated Runtime

Data Size	Cells	Genes	Approximate Time
Small	1,000	2,000	~5 sec
Medium	10,000	5,000	~1-2 min
Large	50,000	10,000	~5-10 min

Times measured with optimized BLAS on Apple M2.

Citation

If you use SCORPION in your research, please cite:

Osorio, D., Capasso, A., Eckhardt, S. G., Giri, U., Somma, A., Pitts, T. M., ... & Kuijjer, M. L. (2023). Population-level comparisons of gene regulatory networks modeled on high-throughput single-cell transcriptomics data. Nature Computational Science. DOI: 10.1038/s43588-023-00558-w