GRNUlar

GRNUlar reconstructs gene regulatory networks from single-cell RNA-sequencing (scRNA-Seq) data using an unrolled deep learning architecture to infer transcription factor (TF)–target interactions.

Key Features:

• Sparse multi-task deep learning with TF priors: Employs a sparse multi-task deep learning framework that incorporates prior information about transcription factors (TFs) to guide GRN reconstruction from scRNA-Seq.
• Unrolled algorithm architecture: Transforms iterative optimization algorithms into neural network layers (unrolled algorithms) to model regulatory relationships and mitigate data noise.
• Supervised training with synthetic scRNA-Seq simulators: Trains the model using synthetic data simulators that generate scRNA-Seq datasets guided by known GRNs for supervised learning.
• GLAD unrolled architecture for undirected GRNs: Applies the GLAD (Graph Learning via Alternating Direction) unrolled architecture to recover undirected GRNs when TF annotations are unavailable.

Scientific Applications:

• GRN reconstruction from scRNA-Seq: Infers TF–target interactions and overall network structure from single-cell RNA-sequencing data.
• Inference with or without TF annotations: Reconstructs directed GRNs when TF identities are provided and recovers undirected GRNs via GLAD when TF information is absent.
• Benchmarking and validation: Applicable to evaluation on synthetic and real datasets for method comparison and network recovery assessment.

Methodology:

Transforms optimization algorithms into unrolled neural network layers, implements a sparse multi-task deep learning model that incorporates TF priors, trains supervisedly on synthetic scRNA-Seq simulators guided by known GRNs, and applies the GLAD (Graph Learning via Alternating Direction) unrolled architecture to recover undirected networks when TF information is missing.