NUTMEG
_{iNsights from single cell mUlTioMics with interpretable GNNs}

NUTMEG is a Python framework for cell-type classification and regulatory program discovery from paired single-cell multiome data, built around graph-based attention.

It brings together data preparation, model training/evaluation, and interpretation into a unified workflow for extracting cell-type-specific regulatory signals.

---

Features

End-to-end multiome workflow

• Data preparation and standardized exports for downstream graph construction and modeling
• Dataset-driven global GRN construction (TF→CRE and CRE→Gene priors)
• Per-cell graph dataset construction
• Model training and evaluation for cell-type classification

Graph-attention modeling

• Two-hop message passing over regulatory graphs (TF → CRE → Gene)
• Graph-based attention weights for edge/path importance

Interpretability & explainability

• Gene contributors per cell type
• Attention-derived cell-type-specific TF–CRE–Gene regulatory programs (eGRNs)

Tutorials

NUTMEG is demonstrated via two notebook tutorials:

1. Data preparation

   • QC, filtering, and dataset exports
   • Notebook: tutorials/01_data_preparation.ipynb

2. End-to-end NUTMEG workflow

   • Build the global GRN
   • Construct the per-cell graph dataset
   • Train + evaluate the model
   • Interpret results
   • Notebook: tutorials/02_nutmeg.ipynb

---

Installation

Clone and install in editable mode:

git clone https://github.com/GeorgakilasLab/NUTMEG.git
cd src/nutmeg
pip install -e .