henhaobio
Sign in

bioRxivSystems BiologyPreprint

cellxgene: a performant, scalable exploration platform for high dimensional sparse matrices

Colin Megill, Bruce Martin, Charlotte Weaver, Sidney M. BelliD, Lia Prins, Seve Badajoz, Brian McCandless, Angela Oliveira PiscoiD, Marcus Kinsella, Fiona GriffiniD, Justin T. KigginsiD, Genevieve Haliburton, Arathi Mani, Matthew Weiden, Madison I. DunitziD, Maximilian Lombardo, Timmy Huang, Trent Smith, Signe Chambers, Jeremy FreemaniD, Jonah CooliD, Ambrose CarriD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Quickly and flexibly exploring high-dimensional datasets, such as scRNAseq data, is underserved but critical for hypothesis generation, dataset annotation, publication, sharing, and community reuse. cellxgene is a highly generalizable, web-based interface for exploring high dimensional datasets along categorical, continuous and spatial dimensions, as well as feature annotation. cellxgene is differentiated by its ability to performantly handle millions of observations, and bridges a critical gap by enabling computational and experimental biologists to iteratively ask questions of private and public datasets. In doing so, cellxgene increases the utility and reusability of datasets across the single-cell ecosystem. The codebase can be accessed at https://github.com/chanzuckerberg/cellxgene. For questions and inquiries, please contact cellxgene@chanzuckerberg.com.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (6 Apr 2021) · CC0 1.0

cellxgene couples a Python server to a browser-based JavaScript client, letting researchers cross-filter, re-embed, annotate, and test differential expression in datasets of millions of cells.

Bioinformaticians explore single-cell data interactively in Jupyter or RStudio, yet passing those insights to experimental biologists and clinicians through static figures and emailed spreadsheets discards context. cellxgene lets experimentalists interrogate and annotate full analysis results themselves, and the authors distinguish it from other visualization tools by biologist-centered design and real-time responsiveness at millions of observations.

Key results

  1. Claimedcellxgene is stated to keep exploration real-time on datasets of millions of cells and tens of thousands of features, relying on D3 WebGL rendering, a purpose-built crossfilter-style filter, and lazy loading of metadata from the server.

    No load-time, latency, or memory benchmarks and no comparison against other viewers are given; scale is stated qualitatively, with the largest dataset at the time of writing described only as millions of cells. · Introduction; Implementation and architecture; Future Directions

  2. ClaimedLinked scatterplot and histogram views support cross-filtering, subsetting, recoloring, re-embedding of subpopulations, on-demand differential expression between arbitrary cell sets, and exportable annotations, e.g. to spot donor-specific clusters signaling uncorrected technical variation.

    Differential expression is computed in-app with t-tests (the stated example) on any user-chosen groups, and coloring by a variable displays its conditional distribution over the others; the workflow is narrated without a worked dataset or statistics. · Core features and functionality; Common uses in single-cell biology

  3. ClaimedAccepts an observation-by-feature matrix, categorical or continuous metadata, and a precomputed low-dimensional or spatial embedding, a format the authors say covers outputs from Seurat, Scanpy, Bioconductor, and scVI; extra results such as clusters import as metadata.

    Inputs listed: a matrix with observations as rows, same-format metadata, and an initial embedding (PCA, tSNE, UMAP, or spatial coordinates); no file format or conversion example is specified, so toolchain compatibility is asserted. · Core features and functionality; Implementation and architecture

  4. Suggestedcellxgene already serves public atlas portals and has been deployed on institutional intranets for private collaborations and inside the Galaxy platform.

    The Figure 1 legend describes cellxgene hosting Tabula Muris Senis, a longitudinal mouse cell atlas; the COVID-19 Cell Atlas and Heart Cell Atlas are named as further deployments; no usage statistics or deployment counts are given. · Implementation and architecture; Figure legends

  5. ClaimedInternal abstractions are domain-agnostic, so any matrix-form data can in principle be loaded; early users have explored single-cell and single-nucleus RNA, ATAC, single-cell spatial, and proteomic data.

    Backed only by a statement of early use across these data types; no example datasets, performance measurements, or details on handling non-transcriptomic features are presented. · Introduction; Implementation and architecture

Methods, data and reproducibility

cellxgene ingests a cell-by-feature matrix, matching metadata, and a precomputed embedding; a Python server lazily serves metadata through a REST API using FlatBuffers serialization, and a JavaScript client renders linked views with D3 WebGL, a crossfilter-inspired filter, and an in-browser dataframe module.

  • Tabula Muris Senis, a longitudinal mouse cell atlas, hosted in cellxgene as the Figure 1 showcase · mouse
  • COVID-19 Cell Atlas web portal, cited as another public cellxgene deployment
  • Heart Cell Atlas web portal, cited as another public cellxgene deployment
Methods & tools
cellxgene (Python server, JavaScript client), REST API with lazy metadata loading, Google FlatBuffers (binary serialization), D3 WebGL, Custom multi-dimensional filter inspired by Mike Bostock's crossfilter, In-browser Pandas-like dataframe module, t-test differential expression, Seurat, Scanpy, Bioconductor, scVI, Galaxy
Reproducibility gaps
The Availability section's link to the hosted tool is an unfilled placeholder; installation instructions are deferred to external documentation, and no tagged release, archived code snapshot, or pinned software versions are cited.
Limitations (4)
  • Per the authors, the largest dataset at the time of writing comprised millions of cells; atlas-scale collections reaching tens of millions of cells will need further work on both system architecture and information display. · stated by the authors
  • At writing, heatmap and dot plot views for comparing cell groups and gene sets and a workspace for gene-level observations were not yet available, and the authors flag in-browser computation as needing improvement. · stated by the authors
  • No load-time, latency, or memory benchmarks, no head-to-head comparison with other single-cell viewers, and no user studies are reported, so the claimed performance and usability advantages cannot be checked from the text. · noted in this summary
  • In-app differential expression is exemplified by t-tests between arbitrary cell selections; with no donor-level modeling or multiple-testing correction described, treating cells as independent samples may overstate significance. · noted in this summary

Read this if you hand single-cell results to wet-lab collaborators and want the design rationale and architecture of a browser viewer meant to stay interactive at millions of cells (no benchmarks included).

Citations
197

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
THISBiology '23Speir '21Goldman '19Li '20Bolisetty '17Ouyang '21
63% semantic matchCell Biology
CZI Single-Cell Biology, S. Abdulla, B. Aevermann+44 more·bioRxiv·Oct 30, 2023