flashscenic

GPU-accelerated SCENIC workflow for gene regulatory network analysis. Seconds instead of hours.

flashscenic replaces the bottleneck steps in the SCENIC pipeline with GPU-powered alternatives: RegDiffusion for GRN inference and vectorized PyTorch implementations of AUCell and cisTarget. The result is a complete GRN analysis pipeline that scales to 20,000 genes and millions of cells, running in seconds on a single GPU.

Installation

pip install flashscenic

For documentation development:

pip install flashscenic[docs]

Requirements: Python 3.9+, PyTorch with CUDA support (CPU fallback available).

Quick Start

We provide a pipeline function run_flashscenic that is capable to cover 90% of use cases.

import flashscenic as fs

# Run the full pipeline in one call
# exp_matrix: (n_cells, n_genes) log-transformed numpy array
# gene_names: list of gene symbols matching columns
result = fs.run_flashscenic(exp_matrix, gene_names, species='human')

# Results
auc_scores = result['auc_scores']       # (n_cells, n_regulons)
regulon_names = result['regulon_names']  # regulon labels

Required resource files (TF lists, ranking databases, motif annotations) are downloaded automatically on first run.

Downloading Data

flashscenic can automatically download the cistarget resource files needed for motif-based pruning:

import flashscenic as fs

# Download human v10 resources (default)
resources = fs.download_data(species='human', version='v10')
print(resources)

# Download mouse resources
resources = fs.download_data(species='mouse')

# List all available resource sets
for rs in fs.list_available_resources():
    print(f"{rs.datasource}/{rs.species}/{rs.version}")

Files are cached in ./flashscenic_data/ by default and skipped on subsequent calls.

Supported species and versions

Step-by-Step Usage

For more control, you can run each pipeline step individually:

import numpy as np
import torch
import flashscenic as fs

# 1. GRN Inference (using RegDiffusion separately)
import regdiffusion as rd
trainer = rd.RegDiffusionTrainer(exp_matrix)
trainer.train()
adj_matrix = trainer.get_adj()

# 2. Filter to known TFs and sparsify
# (load your TF list, subset adj_matrix rows, zero out weak edges)

# 3. Module filtering
filtered_adj = fs.select_topk_targets(adj_matrix, k=50, device='cuda')
filtered_adj, tf_mask = fs.filter_by_min_targets(
    filtered_adj, min_targets=20, min_fraction=0.8
)

# 4. cisTarget pruning
pruner = fs.CisTargetPruner(device='cuda')
pruner.load_database(['db_500bp.feather', 'db_10kb.feather'])
pruner.load_annotations('motifs.tbl', filter_for_annotation=True)
regulons = pruner.prune_modules(modules, tf_names, gene_names)

# 5. AUCell scoring
regulon_adj = fs.regulons_to_adjacency(regulons, gene_names)
auc_scores = fs.get_aucell(exp_matrix, regulon_adj, k=50, device='cuda')

Pipeline Parameters

run_flashscenic exposes all tunable parameters with stage-based prefixes:

Example with custom parameters:

result = fs.run_flashscenic(
    exp_matrix, gene_names,
    species='mouse',
    module_k=100,
    module_min_targets=10,
    module_min_fraction=None,  # disable fraction filter
    pruning_nes_threshold=2.5,
    device='cpu',
)

Core API

Authors

• Hao Zhu (haozhu233@gmail.com)
• Donna Slonim (donna.slonim@tufts.edu)

License

MIT License. See LICENSE for details.