If you’ve worked with single-cell RNAseq data, you’ve probably heard about Seurat. In this blogpost, we’ll cover the the Seurat object structure,in particular the new Seurat version 5. You can also follow the R tutorial here.

So if you are ready… let’s dive in!

Seurat structure overview

Seurat is an R toolkit for single cell genomics. You can read more about it in the Satija lab main page.

When you work with single-cell data in R using the Seurat package, you’ll be working with a Seurat object. What does this mean?

First, let’s talk about what we mean by “single-cell data” and why we need a special type of data structure and not just a “regular ol’ dataframe”.

If we talk about single cell RNA seq data, we’re essentially talking about a table of counts. The rows are genes, and the columns are cells. We can then have metadata, which are annotations about the cells or the genes.  And then we can have more complex data like reductions such as PCA, UMAP or tSNE which reduce the complexity of our dataset to lower dimensions, neighbours… The counts can be also have transformations, sometimes we need to work with raw counts, normalised counts, scaled counts…

This is why we need specific types of objects when we code to store all this data efficiently. Examples are anndata objects if you work in python, SummarisedExperiment or Seurat objects in R.

Seurat object: the “assay” slot

The Seurat object is a representation of single-cell expression data for R. Each Seurat object revolves around a set of cells and consists of one or more assay objects.

The assays have single-cell level expression data (whether that is RNA-seq, ATAC-seq, protein etc). If you are working with single-cell RNAseq data, then the assay will generally be “RNA”, which is the assay we’re going to be focusing on in this blogpost.

Each assay can have several layers. For example, when we first take our scRNAseq sequencing data and create our Seurat object, we will have the create the layer “counts” in the assay “RNA”. Counts is the layer for raw data. But as you very well know, in scRNAseq analysis we often used normalised counts or scaled counts, or batch-corrected counts to compare counts across cells, genes or samples / conditions. These different versions of the raw counts are stored in layers of our Seurat object.

The “data” layer

One of the key steps in data preprocessing involves normalising our data. Normalization is the process of adjusting the raw read counts to account for differences in sequencing depth (i.e., the total number of reads per cell) and other technical factors, ensuring that the expression levels are comparable across all cells. This is important because in scRNA-seq different cells often have different numbers of reads due to various factors like cell size, sequencing efficiency, or differences in the amount of RNA present in each cell.

So for example, if we wanted to compare the expression of Gene B in cells 1 and 2, it would seem like the expression of Gene B is 100 times higher in cell 2 compared to cell 1. But if we check the total number of reads, there were many more reads to begin with in cell 2. We need to correct for this, and that is what normalisation does. The standard way is log2 normalisation, which will make the reads comparable across cells. So now we see that the normalised counts for gene B are the same between cell 1 and cell 2.

The normalised data in the Seurat object is stored in the layer “data”.

The “scale.data” layer

The second key step in data preprocessing is scaling. Scaling adjusts the expression values for each gene so that the expression values across cells are centered and comparable. It is particularly important in analyses like clustering or dimensionality reduction, where you want to ensure that all genes contribute equally to the analysis, rather than being biased by genes with inherently higher expression levels.

Z-scaling means that, for each gene, we subtract the mean expression of that gene across all cells and divide by the standard deviation. This results in a scaled expression value that has a mean of 0 and a standard deviation of 1 across all cells. This process makes the data centered and scaled, so each gene will contribute equally to the analysis.

The scaled data is stored in the layer “scale.data”.

SCTransform: assay “SCT”

If you are used to working with Seurat objects, you might have already heard of SCTransform. SCTransform is an advanced normalization and transformation method specifically designed for single-cell RNA sequencing data. It is an alternative to traditional methods like log-normalization and scaling. The sctransform method uses a statistical model (regularized negative binomial regression) to perform normalization, accounting for both technical noise and biological variation in a more sophisticated way than simpler normalization techniques.

If you are interested in reading more about how sctransform actually works, check out my blogpost on sctransform. You can also check Seurat’s vignette on the SCTransform workflow here.

When we run sctransform on our data, a new “assay” will be created, the SCT assay. Again, you can expect the layers:

• counts, which has “transformed” or corrected raw counts
• data, the log-normalised versions of these corrected counts
• data, which has the Pearson residuals.

Seurat object: the “meta.data” slot

What other information can we store in a Seurat object?

Of course, a very useful slot is the metadata slot. The metadata is used to store information about the cells: whether that is their cell type, their % of mitochondrial counts, whether they were treated or not during an experiment… This enables us, for example, to subset the Seurat object by a particular trait or to remove cells with a particular characteristic, or do differential expression between cells of different cell types. The metadata is just a dataframe with the cell IDs as rows and then different columns with metadata information. Easy!

Seurat object: the “reductions” slot

While processing our single cell data, we often want to visualise our cells. Common dimensionality reduction techniques are PCA, UMAP and t-SNE. We’ll talk about how to actually do this in my blogpost on Seurat workflow, but for now, know that all our dimensionality reduction results will be stored in the slot “reductions”. Examples are pca, tsne or umap. Each reduction stores a dimensionality reduction object, which contains information about the reduced coordinates (e.g., UMAP coordinates) and additional metadata like the variance explained by principal components.

Seurat object: the “graphs” slot

Finally, the graphs slot. It contains a list of graphs (usually a nearest-neighbor graph) that are used for clustering and other analyses. The graph typically represents relationships between cells based on gene expression similarity. For example, RNA_snn (a shared nearest-neighbor graph for RNA-seq data), or SCT_snn. We will talk a bit more about these graphs when we compute them!

Seurat structure: “commands” and “misc” slots

Finally, we’ll quickly mention the slots commands and misc.

Commands contains a record of the commands used to generate the object. This can help in reproducibility by storing a log of the methods and operations that were applied to create or manipulate the Seurat object.

And misc, for miscellaneous, is used to store arbitrary information that doesn’t fit into the other slots. It can be used to store additional analysis results or custom annotations.

Final notes

In a nutshell, a Seurat object is an R S4 object which allows us to store single-cell data in R. 

The key slots in a Seurat object are:

• assays: This slot stores the raw and processed data in different forms. It is a list of Assay objects, each representing a specific type of data.
  • Examples:
    • RNA: The most commonly used assay, containing the raw and processed RNA counts.
    • SCT: Stores data processed using SCTransform (a normalization method).
    • integrated: Contains integrated data when datasets have been merged.
  • Each assay can contain matrices like counts, data, and scale.data.

• meta.data:
  • A data.frame containing metadata associated with each cell. This can include cell type annotations, experimental conditions, or other variables related to the cells.
  • Example columns: cell_type, batch, condition, cluster.

• reductions:
  • A list of dimensionality reductions applied to the data. These are used for visualizations like PCA, t-SNE, or UMAP.
  • Examples: pca, tsne, umap.
  • Each reduction stores a dimensionality reduction object, which contains information about the reduced coordinates (e.g., UMAP coordinates) and additional metadata like the variance explained by principal components.

• graphs:
  • A list of graphs (usually a nearest-neighbor graph) that are used for clustering and other analyses. The graph typically represents relationships between cells based on gene expression similarity.
  • Examples: RNA_snn (a shared nearest-neighbor graph for RNA-seq data), pca_snn.

• clusters:
  • This stores the cluster assignments for each cell after a clustering analysis (e.g., Louvain or Leiden clustering). It is typically stored in the meta.data slot but can also be stored in a separate slot.

• commands:
  • A record of the commands used to generate the object. This can help in reproducibility by storing a log of the methods and operations that were applied to create or manipulate the Seurat object.

• misc:
  • This slot is used to store arbitrary information that doesn’t fit into the other slots. It can be used to store additional analysis results or custom annotations.

Want to know more?

Additional resources

If you would like to know more about Seurat objects, check out:

• Satija lab – Seurat – main webpage

You might be interested in…

• Seurat main workflow and structure in R (R tutorial)

And that is the end of this tutorial!

In this post, I explained how to perform functional enrichment analysis using clusterProfiler. Hope you found it useful!

Before you go, you might want to check:

Ending notes

Wohoo! You made it ’til the end!

In this post, I shared some insights into the Seurat object structure.

Hopefully you found some of my notes and resources useful! Don’t hesitate to leave a comment if there is anything unclear, that you would like explained further, or if you’re looking for more resources on biostatistics! Your feedback is really appreciated and it helps me create more useful content:)

Before you go, you might want to check:

Squids don't care much for coffee,

but Laura loves a hot cup in the morning!

If you like my content, you might consider buying me a coffee.

You can also leave a comment or a 'like' in my posts or Youtube channel, knowing that they're helpful really motivates me to keep going:)

Cheers and have a 'squidtastic' day!