Open-source project
scverse/cellrank avatar
scverse/cellrank

CellRank 2 reads its version from the tag, and its header image has no light mode

CellRank: dynamics from multi-view single-cell data

460 stars58 forksPythonBSD-3-Clause

At a glance

What is it?
scverse/cellrank is a Python framework that turns multi-view single-cell data into fate maps, macrostates and driver genes, built on Markov state modeling with pyGPCCA at its core. One pip line installs it, seventeen required packages come with it, and several small details of the repository do not line up with the README.
Who is it for?
Adopt it if you want a Markov state model with a choice of biological priors rather than one hard wired trajectory method, and if Python 3.12 or newer is already in your environment. Check three things before you commit: that your scanpy, anndata and numpy versions sit above the declared floors, whether the PETSc group behind the fast krylov solver belongs to an extra you can install, and that the submodule declared in .gitmodules is not needed by your workflow.
Can I use it commercially?
Yes. BSD-3-Clause is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 2 days ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • scverse/cellrank460 ★ · +4
Star history
Repository2026-09-112026-11-15
scverse/cellrank456460

4 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The header picture element has a dark source and no img child

The hero graphic at the top of the README is a picture element carrying one source line and nothing else:

html
<picture>
  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/scverse/cellrank/main/docs/_static/img/dark_mode_overview.png">

</picture>

There is no img element inside it and no second source for the light case. The file the source points at is named dark_mode_overview.png, so the single image this block can deliver is the dark variant, and a reader whose system reports no dark color scheme gets nothing, since a picture element with no img child has no fallback of its own. The pair is fenced by image-start and image-end comment markers, and the path is served from the main branch of the repository rather than from a tagged release, so the header image can change under a reader who has already loaded the page.

One pip line, seventeen required packages, and a Python 3.12 floor

Installation is a single command:

bash
pip install cellrank

The dependency list behind it is not small. pyproject.toml requires seventeen packages, each pinned by a lower bound only: anndata>=0.11, docrep>=0.3, joblib>=1.4, matplotlib>=3.9, networkx>=3.4, numba>=0.60, numpy>=2, pandas>=2.2, pygam>=0.10, pygpcca>=1.0.4, rich>=13, scanpy>=1.11, scikit-learn>=1.6, scipy>=1.14, seaborn>=0.13, session-info2>=0.2 and wrapt>=1.16. Every one of those is a floor rather than a pin, so the resolver is free to take the newest release of each, and requires-python is set to >=3.12, with the classifiers naming 3.12, 3.13 and 3.14 and MacOS, Windows and Linux. The README sends anyone who wants other routes to an installation guide, and no requirements file, lock file or conda recipe appears at the top level of the repository.

The version is read from the tag rather than written into pyproject.toml

pyproject.toml declares dynamic = [ "version" ] and names hatchling as the build backend with hatch-vcs in the build requirements, so no version string is written into the file at all. The tag is the only place a release number exists, which makes the tag list load bearing in a way a hard coded version would not be. The recent tags are v2.3.3 on 2026-09-16, v2.3.2 on 2026-06-24 and v2.3.1 on 2026-06-09, and v2.3.3 carries the same timestamp, 09:41:27 UTC, as the last push to the default branch. Tag and commit were recorded in the same second. The README prints no version number anywhere, so a reader comparing a checkout against a PyPI release has nothing in the README to compare.

pyGPCCA is mandatory while its fast krylov solver is not

The backend is named in the README as pyGPCCA, credited to Reuter et al. (2018), and it is a required dependency at pygpcca>=1.0.4 rather than an extra. Two optional dependency groups sit next to it: jax, pulling jax and jaxlib, and moscot, pulling moscot. A comment in pyproject.toml then explains that petsc4py and slepc4py are what give pyGPCCA its fast method='krylov' solver for macrostates. The visible text stops after that sentence, so which optional group the PETSc packages belong to is not stated there. The arrangement to keep in mind is that the numerical core is mandatory, the fast path through it is not, and the extras list is not where you would find the answer.

Badges and the issue link still address theislab, the code sits under scverse

Seven badge links sit above the title and every one of them has empty link text. They cover PyPI, pepy.tech, the GitHub Actions workflow, the readthedocs documentation, codecov, the scverse Discourse forum, and PyPI again: the pypi.org/project/cellrank link appears twice in the same block, once at the top and once at the bottom with five others between them. Three of the destinations name the old organisation path. Actions, codecov and the issue link all address github.com/theislab/cellrank, while the repository itself is scverse/cellrank and the header image is fetched from raw.githubusercontent.com/scverse/cellrank/main. The bug report route a reader is pointed at is the one place the move to the scverse organisation did not reach.

Three agent instruction files and a JavaScript config share a Python root

The top level carries .editorconfig, .git-blame-ignore-revs, .gitmodules, .pre-commit-config.yaml and .readthedocs.yaml alongside AGENTS.md, CLAUDE.md, REVIEW_GUIDE.md, CONTRIBUTING.md and biome.jsonc. Three separate agent instruction files sit in one repository, and biome.jsonc is a JavaScript and TypeScript tool configuration in a package whose primary language is Python, which is a mismatch worth knowing about before a contributor assumes a Node toolchain runs here. The README names none of these files. .gitmodules is the one that would matter most to a user, because a submodule declaration in the tree means something is fetched from elsewhere at clone time, and nothing in the README or the packaging metadata says which paths it covers or whether an install from the published wheel needs any of it.

Five kinds of prior feed one estimator

The application list starts from inputs rather than outputs. Differentiation direction can be estimated from a varied number of biological priors: RNA velocity, citing La Manno et al. (2018) and Bergen et al. (2020), any pseudotime or developmental potential, experimental time points, metabolic labels, and more. From those, the package computes initial, terminal and intermediate macrostates, infers fate probabilities and driver genes, and visualizes and clusters gene expression trends. The design consequence is that the prior is a choice rather than a fixed input, and the README gives no example of switching between them, no guidance on what a disagreement between two priors should look like, and no statement of which combinations are supported.

Two manuscripts, a citation guide, and two related packages with no reference

The README points at a CellRank 1 manuscript and a CellRank 2 manuscript, both in Nature Methods, and then asks readers to follow a citation guide to cite the software correctly, which is a distinction the project takes seriously enough to mark with a warning glyph. The claim that CellRank scales to large cell numbers sits in one sentence with no benchmark, no dataset name and no figure behind it, next to a compatibility claim about the scverse ecosystem and an ease of use claim. The related packages table is more specific: moscot, moslin, VeloVI and RegVelo each carry a reference, while CellMapper and CellAnnotator carry none. Those two are a kNN based label, embedding and molecular layer transfer tool and an LLM based cell type annotator, which makes them the only rows in a table where everything else cites a paper.

Editorial conclusion

Adopt it if you want a Markov state model with a choice of biological priors rather than one hard wired trajectory method, and if Python 3.12 or newer is already in your environment. Check three things before you commit: that your scanpy, anndata and numpy versions sit above the declared floors, whether the PETSc group behind the fast krylov solver belongs to an extra you can install, and that the submodule declared in .gitmodules is not needed by your workflow. Report bugs through the theislab path the README still points at, and keep in mind that no license question arises here since the repository ships BSD-3-Clause with a LICENSE file.

Frequently asked questions

What is CellRank?

CellRank is a modular framework for studying cellular dynamics, built on Markov state modeling of multi-view single-cell data with pyGPCCA at its backend. It installs with one pip line, requires Python 3.12 or newer, and returns macrostates, fate probabilities and driver genes.

How does CellRank compare with scvelo?

This README contains no comparison with scvelo and does not mention that package. What it sets out is CellRank's own set of inputs, among them RNA velocity, any pseudotime or developmental potential, experimental time points and metabolic labels, from which initial, terminal and intermediate macrostates are computed.

Why does installing CellRank pull in scanpy, numba and fifteen other packages?

pyproject.toml lists seventeen required dependencies with lower bounds only, including scanpy>=1.11, numba>=0.60, numpy>=2 and scikit-learn>=1.6. Two optional groups exist as well, jax for jax and jaxlib, and moscot for moscot.

Where should I report a bug in CellRank?

The README links to an issue chooser on github.com/theislab/cellrank, while the repository itself sits under the scverse organisation and its header image is served from raw.githubusercontent.com/scverse/cellrank/main. CONTRIBUTING.md is present at the top level but is not described in the README.

Does CellRank need git submodules to work?

The repository root carries a .gitmodules file, and neither the README nor the packaging metadata says which paths it covers. The published install path is a single pip install cellrank line against the pyproject.toml metadata, which lists seventeen required dependencies and no submodule.

Official sources

  1. License: BSD-3-Clause
  2. Project website
  3. README
  4. Releases
  5. scverse/cellrank on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/scverse-cellrank.svg)](https://hysenlabs.com/projects/scverse-cellrank)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/scverse-cellrank.svg)](https://hysenlabs.com/projects/scverse-cellrank)

Was this useful?