NETMAP Documentation

NETMAP infers single-cell gene regulatory networks (GRNs) using explainable AI. An ensemble of count-distribution autoencoders is trained on scRNA-seq data, then Captum attribution methods score directed gene–gene influences per cell. The full attribution matrix (cells × genes²) is stored as an AnnData object.

Pipeline overview

  1. Model training — netmap.model trains an ensemble of ZINBAutoencoder or NegativeBinomialAutoencoder models with early stopping via create_model_zoo().

  2. GRN inference — inferrence() applies Captum attribution (GradientShap, GuidedBackprop, or Deconvolution) across all model–target combinations and writes Parquet-backed output.

  3. Masking — netmap.masking annotates edges with co-expression support (internal) and external reference GRN overlap (external).

  4. Downstream — netmap.downstream clusters the GRN, ranks edges, builds signed regulons, and produces interactive network visualisations.

Installation

# Recommended — includes GPU deps (PyTorch, CuPy), CUDA 12.8, linux-64 only
pixi run install-kernel

# pip — GPU dependencies must be added manually
pip install -e .

Tutorials

API Reference