SCENIC

SCENIC begins by identifying transcription factor (TF)–target gene co-expression modules using GENIE3 or GRNBoost2, gradient-boosting regression algorithms that predict TF expression from target gene expression. This step captures potential regulatory relationships based on correlated expression patterns across thousands of single cells, generating an initial adjacency matrix of TF–gene interactions.

Each co-expression module undergoes RcisTarget analysis, which scans gene promoters and enhancers for enriched cis-regulatory motifs—short DNA sequences recognized by transcription factors. Only modules where the TF's binding motif is significantly overrepresented in target gene regulatory regions are retained, filtering out indirect correlations and false positives to produce direct regulatory links.

SCENIC calculates AUCell scores for each regulon in every cell by evaluating the enrichment of regulon target genes within the cell's expressed gene set. This produces a binary activity matrix indicating which regulons are active in which cells, transforming raw expression data into a functional readout of transcription factor activity suitable for downstream clustering and trajectory analysis.

Cells are clustered using regulon activity scores rather than raw gene expression, grouping cells by shared regulatory programs instead of transcriptional similarity alone. This approach reveals biologically meaningful populations that may be obscured in expression-based clustering, such as rare progenitor states defined by specific TF activity signatures.

The Python reimplementation, pySCENIC, provides a scalable, containerized workflow with improved performance for large datasets. Key enhancements include:

• Distributed execution via Dask for parallel motif enrichment
• Feather file format for efficient intermediate storage
• Nextflow pipeline support for reproducible, automated runs
• Compatibility with Scanpy and AnnData ecosystems

SCENIC supports regulatory network inference across multiple species through configurable motif databases and reference genomes. Built-in databases include human (hg38), mouse (mm10), and Drosophila (dm6), with the ability to supply custom motif collections for non-model organisms. This flexibility enables comparative regulatory biology studies and translational research applications.

The computational reconstruction of transcription factor–target gene interactions from expression data. SCENIC is a leading method in this field, moving beyond simple correlation by validating predicted targets with cis-regulatory motif enrichment. Key steps include:

• Identifying co-expression modules
• Pruning targets using motif analysis
• Scoring regulon activity per cell

A regulon is the fundamental unit of SCENIC's output, defined as a transcription factor and its set of directly regulated target genes. Unlike gene sets, regulons are validated by the presence of the transcription factor's binding motif in the target genes' regulatory regions. Regulon activity scores, measured by Area Under the Curve (AUC), provide a robust, binarized readout of transcription factor activity in each cell.

The computational process that distinguishes SCENIC from pure co-expression networks. Tools like RcisTarget scan the regulatory sequences of co-expressed gene sets to identify statistically overrepresented transcription factor binding motifs. This step filters out indirect correlations and false positives, ensuring that only genes with a high likelihood of direct physical regulation by the transcription factor are retained in the final regulon.

The algorithm used to score regulon activity in individual cells. AUCell calculates the Area Under the Curve for the recovery of a regulon's target genes within the ranked gene expression profile of each cell. This generates a continuous activity score that is then binarized into an 'on' or 'off' state, enabling the identification of which transcription factor-driven programs are active in specific cell populations.

A complementary method that predicts the future transcriptional state of individual cells by modeling the ratio of unspliced to spliced mRNA. While SCENIC identifies the static regulatory logic of a cell's current state, RNA velocity reveals the direction and speed of state transitions. Integrating regulon activity with velocity vectors can pinpoint the transcription factors driving differentiation trajectories.

Large-scale pretrained models like Geneformer and scGPT learn universal cell representations from massive single-cell corpora. These models can be fine-tuned to predict transcription factor targets or gene expression responses to perturbations. They represent a deep learning evolution of the network inference problem that SCENIC solves, offering context-aware predictions without explicit motif analysis.