Seurat v4 vs Seurat v5

Seurat versions comparison

Here I summarize the most relevant conceptual and structural changes introduced in Seurat v5 compared to Seurat v4, with an emphasis on how these changes impact single-cell RNA-seq analysis workflows.

This notebook is intended to frame the new conceptual model behind Seurat v5 and its alignment with current best practices.

Versioning note
This overview reflects Seurat v5.x behavior as of early 2026. Minor interface details may evolve across releases, but the core design principles described here are stable.

Unless explicitly stated otherwise, examples and concepts discussed here refer primarily to RNA assays. Multimodal (RNA + ATAC) considerations are noted where relevant.

What does Seurat v5 introduce?

Seurat v5 introduces an internal reorganization of the data object and the analysis workflow with the following goals:

  • Improve scalability for large datasets
  • Facilitate multimodal analyses
  • Allow methodological flexibility in normalization and integration
  • Reduce unnecessary data duplication within the object

These changes reflect a shift toward more modular, explicit, and comparable workflows.

Main structural change: from ´Assay´ to ´Assay5´

Seurat v4

  • Each Assay contains fixed slots:
    • counts
    • data
    • scale.data
  • A single normalization strategy per assay

Seurat v5

  • Introduces Assay5 and the concept of layers
  • Allows multiple representations of the same data within a single assay
  • Explicit separation between:
    • raw data
    • normalized data
    • transformed data

This design avoids object duplication and enables alternative analytical strategies without recomputing the entire workflow.

The concept of layers in Seurat v5

A layer represents a specific data representation within an assay. Common examples include:

  • counts
  • lognorm
  • sctransform
  • pearson_residuals

This structure provide us a key of advantages, now we can:

  • Compare different normalization methods within the same object
  • Modify downstream analyses without recomputing upstream steps
  • Enable methodological benchmarking

Normalization: conceptual differences between v4 and v5

Seurat v4

  • Common workflows use NormalizeData() (methods: LogNormalize, CLR, RC) or SCTransform() as a separate variance-stabilizing normalization framework
  • Normalization choice is typically exclusive and assay-bound

Seurat v5

  • The layer-based data model allows multiple normalized representations to coexist within the same assay
  • External normalization strategies (e.g. scran, shifted-log) can be stored as layers
  • Pearson residuals are supported as first-class outputs in residual-based workflows
  • Downstream analyses explicitly reference the layer being used, rather than inheriting assumptions from the assay

Additional workflow-relevant changes

  • Feature selection
    Expanded from predominantly variance-based HVGs toward analytical, model-based approaches such as Pearson residuals and Deviance, while retaining variance-based methods.

  • Dimensionality reduction
    Decoupled from the assay and explicitly defined by layer, improving clarity and flexibility.

  • Clustering
    Shift from Louvain algorithm to Leiden as the default method, althought Louvain is still supported. This change push to get a better partition stability and faster convergence.

    Related resources reviewed on 2025, but still on practice:

Ver presentación: Metodos de agrupamiento en datos multiome

  • Differential analysis
    Growing emphasis on pseudobulk and donor-aware modeling, aligned with recent benchmarking studies, while retaining cell-level testing options

Summary comparison

Aspect Seurat v4 Seurat v5
Data object Fixed-slot Assay Assay5 with layers
Data representations One per assay Multiple layers
Normalization LogNormalize or SCTransform Multiple stored representations (e.g. scran, residuals)
Feature selection Variance-based HVGs Pearson residuals and Deviance
Dimensionality reduction Assay-coupled Layer-defined
Default clustering Louvain Leiden
DGE Mostly cell-level Pseudobulk and mixed models

What changes if I migrate today?

Aspect If you stay on Seurat v4 If you migrate to Seurat v5
Object structure Fixed slots per assay Assay5 with multiple layers
Normalization One method per assay Multiple normalization strategies within the same assay
Method comparison Requires recomputation Native support via layers
Clustering default Louvain Leiden

Addendum: Signac workflows (RNA + ATAC) in Seurat v5

Seurat multiome

While Seurat v5 provides a unified object model for multimodal data, Signac-based chromatin workflows introduce additional constraints that are important to keep in mind during migration.

Key considerations:

  • Chromatin assays are more sensitive to object conversion than RNA assays
    In particular, legacy objects created under Seurat v4 + Signac v1.x may not always retain full chromatin functionality when updated to Seurat v5.

  • Assay5 does not automatically guarantee chromatin compatibility
    A converted chromatin assay may exist structurally but fail downstream operations (e.g. peak-based reductions, motif analysis, or linkage steps).

  • Recommended practice
    For multimodal projects (RNA + ATAC), especially those involving:
    • peak calling
    • TF motif analysis
    • peak–gene linkage

    it is often safer to reconstruct the chromatin assay directly under Seurat v5 / Signac v2, rather than relying on object conversion.

  • RNA assays are generally robust to conversion, while ATAC assays benefit from explicit reprocessing.

This distinction is particularly relevant for large multiome datasets and for workflows that rely heavily on chromatin-specific reductions and annotations.

Alignment with Seurat v5.1+ developments

The conceptual framework described in this notebook remains stable across Seurat v5 releases. However, recent versions (v5.1+) further reinforce the design principles introduced in v5.0.

These refinements include:

  • More consistent handling of layers across functions
  • Improved interoperability with external methods
  • Clearer separation between data representation and analysis logic

I have found that Seurat v5.1+ have make a clear improvement toward transparent, modular, and benchmark-aligned workflows, rather than introducing breaking conceptual changes.

Users migrating from Seurat v4 are therefore encouraged to adopt the layer-centric mindset early, as it represents the stable direction of the framework going forward.

This notebook serves as a conceptual reference to understand these changes and to support informed decisions when designing new analyses or migrating existing projects.

Resources

  1. Seurat v5 announcements and design overview

  2. Seurat v5 Command Cheat Sheet (Assay5 + layers)

  3. Hafemeister & Satija (2019) – SCTransform

  4. Crowell et al. (2020) – Benchmarking scRNA-seq analysis pipelines

  5. Squair et al. (2021) – Best practices for differential expression in scRNA-seq

  6. Traag et al. (2019) – Leiden clustering algorithm

CSC. February 2026

This site uses Just the Docs, a documentation theme for Jekyll.