cellink: a framework for joint analysis of genotype and single-cell data

Single-cell profiles are indexed by cell; genotype data is indexed by donor. Keeping the two correctly paired through subsetting and filtering is easy to get wrong by hand, and each downstream genetics tool (PLINK, MAGMA, LDSC, TensorQTL, SAIGE-QTL, and more) expects its own file formats and conventions. cellink provides a single DonorData structure that keeps donor- and cell-level data synchronized, and the glue to drive that surrounding tool ecosystem directly from it.

Key Features

1. Unified DonorData structure

cellink introduces the DonorData class, unifying individual-level and single-cell data. It extends standard formats (AnnData, MuData) with GenoAnnData for efficient genotype (via dask) and phenotype (via ehrapy) handling.

• Donor-level Data (G): GenoAnnData, Stores individual level data such as genotypes.
• Cell-level Data (C): AnnData/ MuData, Stores single-cell omics data such as gene expression.

Crucially, DonorData ensures that genetic data and single-cell modalities remain synchronized, preserving their donor-cell pairing even through complex filtering operations (e.g., selecting specific cell types or patient subsets). See the DonorData on-disk format for how this is represented on disk and how to check a DonorData object against it.

2. Comprehensive toolkit

cellink offers a streamlined suite of tools for the entire analysis workflow, organized by task:

Association testing

• eQTL analysis with jaxQTL or tensorQTL
• eQTL analysis with SAIGE-QTL
• Naive pseudobulk eQTL mapping
• Rare variant association studies
• Clumping & pruning
• Colocalization analysis
• Resolving association-test inputs directly from DonorData via formula strings

Heritability, gene programs & GWAS integration

• Cell-type specific LD score regression
• Gene program scoring with sc-linker
• Gene-set enrichment with MAGMA
• Integrating GWAS with single-cell disease relevance scores (scDRS/seismic)
• Integrating GWAS with spatial data (gsMap)

Deep learning & representation learning

• LIVI: donor-level representation learning
• Scooby: single-cell-resolution sequence-to-coverage modeling & variant scoring
• Built-in dataloaders for deep learning, including Multiple Instance Learning (MIL) over per-donor cell bags

Other

• Variant preprocessing & annotation: quality control, annotation (VCF export/import), and selection of genetic variants.
• Integrating DonorData with EHR data

Getting Started

Install the latest development version directly from GitHub (cellink is not yet published on PyPI; a future release will be published as cellink-tools, see the installation guide for optional extras and why):

pip install git+https://github.com/theislab/cellink.git@main

New to cellink? Start with DonorData basics: no analysis, just how to build a DonorData from your own genotype/expression data, slice it, and save it. From there, the Documentation section below is the map to everything else.

Documentation

Installation guide	Requirements and every optional extra (pip install cellink[extra-name])
Tutorials	Step-by-step guides for every analysis workflow above
API reference	Every public function and class, organized by module
DonorData on-disk format	The versioned HDF5/Zarr schema, and how to check an object against it
Contributing guide	Dev setup, tests, and the PR workflow
Changelog	What changed in each release

Contact

If you found a bug, please use the issue tracker.

Release notes

t.b.a

Citation

t.b.a