Reference implementation

Code

The PyTorch implementation is released through the GAHIB repository with a Scanpy-style workflow for single-cell AnnData inputs.

Open GitHub repository

Install

Editable package

pip install -e .
# optional extras
pip install -e ".[all]"
pip install -e ".[dev]"

Python 3.9 or later. Full dependency list is maintained in the repository package metadata.

Outputs

Model accessors

model.get_latent() returns the learned embedding.
model.get_time() returns the trajectory time estimate when trajectory modules are enabled.
model.get_centroid() returns per-cluster latent centroids.

Quick start

Fit GAHIB on an AnnData object

The graph encoder is selected explicitly; set encoder_type to mlp for a plain VAE without graph message passing.

import scanpy as sc
from gahib import GAHIB

adata = sc.read_h5ad("data/human_cd34_bone_marrow.h5ad")

model = GAHIB(
  adata,
  layer="counts",
  encoder_type="graph",
  graph_type="GAT",
  latent_dim=10,
  i_dim=2,
  irecon=1.0,
  lorentz=5.0,
)

model.fit(epochs=200, patience=30)
Z = model.get_latent()

Notebooks

A Colab demo notebook with a frozen checkpoint will be released alongside the paper. Subscribe to the GitHub repository for the announcement.

License

The reference implementation is distributed under the MIT License.

Continue

Model

Method

Review the architecture before running experiments.

Open route ->Inputs

Datasets

Match the public metadata against your local data layout.

Open route ->Credit

Cite

Copy the current citation stub and publication-state notes.

Open route ->