deepSCENIC

Deep learning for single-cell Gene Regulatory Networks

deepSCENIC learns hierarchical TF→region→gene regulatory cascades by integrating scRNA-seq, scATAC-seq, and DNA sequence data. It combines a VAE-based multimodal framework with Enformer-derived sequence embeddings to infer cell-type-specific gene regulatory networks.

Key Features

• Multimodal integration: Jointly models RNA expression and chromatin accessibility
• Sequence-informed: Uses Enformer to learn TF binding patterns from DNA sequence
• Perturbation prediction: Simulates TF knockdown/overexpression effects
• scverse compatible: Works with AnnData, MuData, and scanpy workflows

Installation

pip install deepscenic

If you want to use the enformer model as the sequence model (this is the default):

pip install deepscenic[enformer]

Requirements: Python ≥3.12, PyTorch ≥2.6

Documentation

Full documentation including tutorials and API reference: deepscenic.readthedocs.io

Quick Start

deepSCENIC follows the scanpy-style API with modules for preprocessing (ds.pp), tools (ds.tl), and plotting (ds.pl).

Data Preparation

import deepscenic as ds

# Annotate transcription factors on the RNA data
tfs = ds.fetch_tf_collection(species="human")
ds.pp.mark_tfs(adata_rna, tf_list=tfs)

# Add gene annotations (for chromosome-based splitting)
annot, chromsizes = ds.fetch_gene_annotation(species="hsapiens")
ds.pp.add_gene_annotation(adata_rna, gene_annotation=annot)

mdata = ds.pp.create_mudata(rna=adata_rna, atac=adata_atac)

# Compute region-to-gene search space (regions within 1Mb of gene TSS)
ds.pp.compute_r2g_penalty(mdata)

# Split cells and features for training/evaluation
ds.pp.split_cells(mdata, test_fraction=0.2, stratify_key="cell_type")
ds.pp.split_features_by_chromosome(mdata, test_chromosomes=["chr7", "chr11"])

Model Training

# Register genome for sequence extraction
ds.register_genome("/path/to/hg38.fa")

# Phase 1: Train full model (VAE + MotifNet + Enformer)
model = ds.tl.train(mdata, epochs=100, device="cuda")

# Phase 2: Recompute tf2r and finetune all r2g links
model = ds.tl.finetune_r2g(
    model,
    mdata,
    epochs=500,
)

# Save trained model
model.save("my_model.pt")

GRN Extraction & Analysis

# Extract gene regulatory network
grn = ds.tl.extract_grn(model)

# Get targets of a specific TF
targets = ds.tl.get_tf_targets(model, tf_name="SOX10", top_k=100)

# Visualize GRN heatmap
ds.pl.heatmap_grn(model, tfs=["SOX10", "MITF", "PAX3"])

Perturbation Simulation

# Simulate TF knockdown (level=0) or overexpression (level=2)
perturbed, logFC = ds.tl.simulate_perturbation(
    model,
    mdata,
    tf_name="SOX10",
    level=0,  # knockdown
)

# Summarize and visualize perturbation effects
results = ds.tl.process_perturbation_results(logFC, mdata, tf_name="SOX10")
ds.pl.waterfall_perturbation(results)

For complete workflows, see the tutorials.

Citation

If you use deepSCENIC in your research, please cite:

Partel G, et al. DeepSCENIC: transfer learning from sequence-to-function models enables causal gene regulatory network inference. bioRxiv (2026). doi:10.64898/2026.09.18.752607

@article{partel2026deepscenic,
  title={DeepSCENIC: transfer learning from sequence-to-function models enables causal gene regulatory network inference},
  author={Partel, Gabriele and De Winter, Seppe and Konstantakos, Vasileios and Blaauw, Casper H. and Aerts, Stein},
  journal={bioRxiv},
  year={2026},
  doi={10.64898/2026.09.18.752607}
}

Links

• Documentation
• GitHub Repository
• Issue Tracker
• Enformer PyTorch - Sequence model implementation