LoginSign Up
← Back to all sparks
S

Seurat

ANALYTICS
Velocity0.0

Toolkit for quality control, analysis and exploration of single-cell RNA-seq data

github.com/satijalab/seurat ↗

Seurat's centre of gravity has moved from single cells to spatial data and on-disk matrices

single-cellspatial-transcriptomicsbioinformaticson-disk-matricesvisualization
◆Current state
Seurat is the dominant R toolkit for single-cell analysis, and the 5.x line reads as two ongoing projects. One is spatial: successive releases absorb each new 10x output format - Visium HD, Xenium protein data, Space Ranger 4.0 segmentations - and add plotting and selection tools for them. The other is scale, where BPCells on-disk matrices keep gaining support in functions that previously required everything in memory.
◆Where it's heading
Both projects are driven from outside. The spatial work tracks whatever 10x ships, which is why data loaders and coordinate handling get rewritten release after release; the BPCells work tracks dataset sizes that no longer fit in RAM. Clustering and dimensionality reduction, the parts Seurat actually owns, change mainly by exposing more of uwot's and igraph's options rather than by new method development.
◆Prediction
Expect the next release to absorb whatever instrument output 10x publishes next, and BPCells support to keep spreading into the functions that still densify matrices; the interactive spatial selection tooling looks like the one area with room to grow on its own terms.

◆Recent moves

  1. 1mo ago

    Fixes zero-count genes in PrepSCTFindMarkers; UMAP init option

    Mostly correctness, and some of it consequential: an incorrect theta filter was producing zero counts for valid genes in multi-model PrepSCTFindMarkers, and supplied features could be dropped from RunPCA on v5 objects when unrelated features had zero variance. Both are the sort of bug that silently changes results rather than erroring.

    View source ↗
  2. 3mo ago

    BPCells gains regression support; ElbowPlot shows variance explained

    The on-disk thread advances: ScaleData on a BPCells matrix now supports vars.to.regress and latent.data, and FindMarkers warns when handed column-major BPCells storage. ElbowPlot can finally plot percent and cumulative percent variance explained rather than only standard deviation.

    View source ↗
  3. 8mo ago

    Space Ranger 4.0 segmentations and interactive cell lasso

    ⚡ SPARK

    The spatial thread's clearest step: Seurat absorbs 10x's segmented Visium output across its loaders, coordinate handling and every spatial plotting function, and adds interactive lasso selection of cells. The coordinate fix in the same release - x now maps to imagecol as 10x defines it - shows how much of this is conforming to a vendor's model.

    View source ↗
  4. 9mo ago

    Xenium protein data; Leiden via igraph, UMAP via umap2

    More instrument coverage with LoadXenium reading protein data, alongside a pattern that recurs throughout this timeline: rather than implement methods, Seurat exposes the upstream ones - cluster_leiden from igraph, umap2 and approx_pow from uwot - as options.

    View source ↗
  5. 1y ago

    sctransform and leverage-score calculation refactored for speed

    A performance-only release refactoring sctransform and leverage score calculation, with edge-case fixes in variable feature selection and module scores. It targets the same problem BPCells does - datasets that have outgrown the original implementation.

    View source ↗
  6. 1y ago

    Test fix for cross-platform clustering variability

    A same-day follow-up to 5.2.0 adjusting a test to tolerate FindClusters assigning labels differently across systems. Nothing changes for users.

    View source ↗