torch-pba 0.0.3

PyTorch implementation of PBA.

torch-pba

PyTorch Implementation of PBA. AnnData-centric.

Installation

From PYPI:

pip install torch-pba

Alternatively, install the development version from GitHub:

git clone https://github.com/mvinyard/torch-pba.git; cd torch-pba; pip install -e .

Example use:

from torch_pba import PBA
from anndata import read_h5ad

pba = PBA(adata=read_h5ad("./path/to/adata.h5ad"))

pba.build_kNN()
pba.compute_Laplacian()
pba.compute_potential()
pba.compute_fate_bias()
pba.compute_mean_first_passage_time()

Time to calculate Mean First Passage Time for the example hematopoiesis dataset is cut from 4+ hours to <10 mins. In this example, I used a NVIDIA T4 GPU rented from GCP.

See more: notebook

Original work:

• GitHub: AllonKleinLab/PBA
• Paper: Weinreb et al., PNAS. DOI: 10.1073/pnas.1714723115. (2018).

Note:

I have not contributed any methodological novelty in this library. The original implementation contains the novel application of a Laplace transform to a kNN Graph to obtain a potential value, pseudotime, etc. Here, I have simply adapted the library to PyTorch/CUDA. No formal benchmarking has been performed.

Contact / questions:

[e-mail]