henhaobio
Sign in

Genome biologyBioinformaticsPeer-reviewed journal article

Scanpy for analysis of large-scale single-cell gene expression data

F. Alexander WolfiD, Philipp AngereriD, Fabian Joachim TheisiD

PDFDOIPreprintCode

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text (9 Aug 2017) · CC0 1.0

Python toolkit Scanpy, built on HDF5-backed AnnData, achieves speedups of 5 to 16 times over Cell Ranger's R kit and clusters 1.3 million cells without subsampling.

Integrated single-cell workflows in frameworks such as Seurat, Monocle, MAST, and Cell Ranger do not scale to data sets of one million cells or more, and these R-based tools are harder to couple with Python machine-learning libraries. Scanpy supplies comparable preprocessing, visualization, clustering, trajectory, and differential expression analyses in Python around a sparse, disk-backable data class, extending them to million-cell data.

Key results

  1. ShownMillion-cell data can be analyzed in full on modest hardware: 1.3 million cells were visualized and clustered without any subsampling.

    1.3 million E18 mouse brain cells from 10x Genomics were processed in a few hours on eight cores of a small server (Fig. 1c); exact run time and peak memory are not reported. · SCANPY is benchmarked in comparisons with…; Figure legends

  2. ShownOn 68,579 peripheral blood mononuclear cells (PBMCs), Scanpy ran representative analysis steps markedly faster than the run-time-optimized Cell Ranger R kit.

    Speedups of 5 to 16 times over the Cell Ranger R kit across the representative steps compared in Fig. 1b; only this range is given in the text, with no per-step timings or hardware description. · SCANPY is benchmarked in comparisons with…; Figure legends

  3. ShownReproducing a Seurat clustering tutorial end to end, from raw counts to cell-type identification, Scanpy was faster than Seurat at every step.

    Per-step speedups ranged between 5 and 90 times on 2700 PBMCs; the step-by-step comparison lives in the scanpy_usage repository (170505_seurat) rather than in a table in the text. · SCANPY is benchmarked in comparisons with…; Availability and requirements

  4. ShownApplied to the 68,579 PBMCs, the workflow (confounder regression, normalization, highly variable genes, Louvain clustering, marker ranking, pseudotime) produced a cluster marker matching bulk-derived labels and a branching trajectory.

    Differential expression ranking singled out MS4A1 as the B-cell marker of cluster 7, consistent with labels from comparisons with bulk expression; diffusion pseudotime rooted at a cell in the CD34+ cluster revealed a branch, displayed on tSNE and diffusion maps (Fig. 1a). · Figure legends; SCANPY integrates canonical analysis methods in…

  5. ClaimedScanpy's efficiency is attributed to design choices: a sparse, HDF5-backable AnnData container usable without full loading into memory, in-place operations, parallelized tools, and a single reusable neighborhood graph computed faster than in a popular reference package.

    Neighbor distances are computed per row-aggregated submatrix using parallelized matrix multiplication; the text gives no timings, memory figures, or name of the reference package for the graph and backed-mode claims. · SCANPY introduces efficient modular…; Comparison with existing Python packages for…

Methods, data and reproducibility

Scanpy implements preprocessing, tSNE, graph-drawing and diffusion-map visualization, graph-based Louvain clustering, differential expression ranking, diffusion pseudotime, and gene regulatory network simulation around AnnData, an HDF5-backable annotated matrix class; run time is benchmarked against Seurat and the Cell Ranger R kit on 10x Genomics data.

  • Peripheral blood mononuclear cells analyzed for the Cell Ranger R kit speed comparison and the Fig. 1a workflow demonstration · n = 68,579 cells
  • PBMCs from a Seurat clustering tutorial, reanalyzed from raw counts to cell types for the Seurat speed benchmark · n = 2700 cells
  • E18 mouse brain cells from 10x Genomics, visualized and clustered without subsampling · mouse · n = 1.3 million cells
  • Data from the original diffusion pseudotime study, reanalyzed to reconstruct branching developmental processes (scanpy_usage 170502_haghverdi16)
  • Single cells simulated from literature-curated gene regulatory networks (scanpy_usage 170430_krumsiek11)
  • Deep-learning results on single-cell imaging data (scanpy_usage 170529_images)
Methods & tools
Scanpy (Python toolkit for single-cell expression analysis), AnnData (annotated data matrix class with sparse and HDF5-backed storage), Seurat (speed comparator via its 2700-PBMC clustering tutorial), Cell Ranger R kit (speed comparator on 68,579 PBMCs), Louvain clustering, Diffusion maps and diffusion pseudotime, tSNE and Fruchterman-Reingold graph drawing, Principal component analysis, NumPy, SciPy, pandas, Matplotlib, h5py (core dependencies), scikit-learn, statsmodels, seaborn, NetworkX, igraph (used by parts of the toolkit), loompy / loom format (AnnData import and export)
Software & environment
Python package on PyPI (BSD3 license) for Linux, Mac OS, and Windows; core depends on NumPy, SciPy, Matplotlib, pandas, h5py, with scikit-learn, statsmodels, seaborn, NetworkX, igraph, tSNE and Louvain packages for some tools; no versions given.
Assay conditions & compute
The 1.3 million-cell analysis took a few hours on eight cores of a small server; interactive run times of seconds are stated for about 100,000 cells; CPU model, memory, and benchmark hardware are not given.
Reproducibility gaps
No software versions, hardware specifications, per-step timings, or memory measurements appear in the text; benchmark details are deferred to external scanpy_usage notebooks, and the three 10x Genomics data sets are cited without URLs or accessions.
Limitations (3)
  • The Python ecosystem underlying Scanpy offers fewer options for classical statistical analyses than R. · stated by the authors
  • Speed benchmarks give only ranges (5 to 16 and 5 to 90 times) without per-step timings, hardware, or memory use, and the text reports no concordance of clusters or markers with Seurat or Cell Ranger outputs. · noted in this summary
  • The 1.3 million-cell demonstration covers only visualization and clustering, and the faster neighborhood-graph computation and HDF5 backed mode are asserted without quantitative benchmarks or a named reference package. · noted in this summary

Read this if you analyze large single-cell expression data sets, up to a million cells or more, in Python, or want the design rationale behind AnnData and Scanpy's shared neighborhood graph.

Citations
8,891
FWCI
121.92
Field percentile
top 1%

Abstract

We present Scanpy, a scalable toolkit for analyzing single-cell gene expression data. It includes preprocessing, visualization, clustering, pseudotime and trajectory inference, differential expression testing and simulation of gene regulatory networks. The Python-based implementation efficiently deals with datasets of more than one million cells and enables easy interfacing of advanced machine learning packages. Code is available from https://github.com/theislab/scanpy.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGardeux '17Zhang '20Moreno '21Xuan '26Abdelaal '20Zhang '22
65% semantic matchBioinformatics
Vincent Gardeux, Fabrice P.A. David, Adrian Shajkofci+2 more·Bioinformatics·May 23, 2017· 134 citations