Single-Cell Omics

The manylatents.singlecell module provides data loading for single-cell omics datasets stored in the AnnData .h5ad format, covering scRNA-seq, scATAC-seq, and CITE-seq assays.

---

Overview

Single-cell experiments measure gene expression (or chromatin accessibility, surface proteins, etc.) in individual cells. The resulting count matrices are stored as AnnData objects in .h5ad files, which manylatents-omics loads directly into PyTorch for dimensionality reduction and geometric analysis.

Install: uv add "manylatents-omics[singlecell]"

---

Shipped Datasets

Preconfigured Hydra configs are provided for common benchmark datasets:

Dataset	Cells	Features	Config
PBMC 3k	~2,700	~1,800 genes	data=pbmc_3k
PBMC 10k	~10,000	varies	data=pbmc_10k
PBMC 68k	~68,000	varies	data=pbmc_68k
Embryoid Body	varies	varies	data=embryoid_body

---

Key Classes

AnnDataset

manylatents.singlecell.data.AnnDataset

A PyTorch Dataset for any AnnData .h5ad file. Supports:

• Loading from adata.X, adata.raw.X, or a named layer (adata.layers[layer])
• Automatic sparse-to-dense conversion
• Cell-type label extraction from adata.obs with integer encoding
• Access to observation annotations via get_obs(key)

AnnDataModule

manylatents.singlecell.data.AnnDataModule

A PyTorch Lightning DataModule wrapping AnnDataset. Supports two modes:

• full: Entire dataset used for both training and testing
• split: Random train/test split with configurable ratio and seed

---

Usage

# UMAP on PBMC 3k
python -m manylatents.main data=pbmc_3k algorithms/latent=umap

# Sweep datasets and algorithms
python -m manylatents.main -m \
  data=pbmc_3k,pbmc_10k \
  algorithms/latent=umap,phate

Loading a Custom Dataset

To use your own .h5ad file, create a Hydra config or instantiate the datamodule directly:

from manylatents.singlecell.data import AnnDataModule

dm = AnnDataModule(
    adata_path="path/to/your_data.h5ad",
    label_key="cell_type",
    batch_size=128,
    mode="full",
)
dm.setup()