Background

The tomo-seq technique is based on cryosectioning of tissue and performing RNA-seq on consecutive sections. Unlike common RNA-seq which is performed on independent samples, tomo-seq is performed on consecutive sections from one sample. Therefore tomo-seq data contain spatial information of transcriptome, and it is a good approach to examine gene expression change across an anatomic region.

This vignette will demonstrate the workflow to analyze and visualize tomo-seq data using tomoda. The main purpose of the package it to find anatomic zones with similar transcriptional profiles and spatially expressed genes in a tomo-seq sample. Several visualization functions create easy-to-modify plots are available to help users do this.

At the beginning, we load necessary libraries.

library(SummarizedExperiment)
library(tomoda)

Dataset

This package contains an examplary dataset geneated from 3 day post cryinjury heart of zebrafish, obtained from GSE74652. The dataset contains the raw read count of 16495 genes across 40 sections. Here we load the dataset and view the first several rows of it.

data(zh.data)
head(zh.data)
#>                    X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17
#> ENSDARG00000000002  1  0  0  0  0  2  0  3  1   1   3   0   0   4   2   7   3
#> ENSDARG00000000018  0  0  0  0  0  2  2  4  1   6   3   2   2   6   1   3   1
#> ENSDARG00000000019  4  0  1  2  1  0  4  1  4   0   6   9   2   9   1   8   3
#> ENSDARG00000000068  1  0  0  0  0  0  2  4  2   1   3   0   1   1   1   2   0
#> ENSDARG00000000069 13  0  1  0  0  1  5  4  5   7  14   8   3   8   2   8  10
#> ENSDARG00000000086  0  0  0  0  0  0  0  0  0   0   0   1   1   2   0   1   0
#>                    X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31 X32
#> ENSDARG00000000002   3   2   0   7   4   1   0   0   1   3   0   0   0   0   3
#> ENSDARG00000000018   0   1   1   2   1   5   1   0   2   1   5   0   0   0   2
#> ENSDARG00000000019  11   6   2   9   4  12   1   2   6   9   1   4   4   5   7
#> ENSDARG00000000068   0   0   0   0   2   1   0   1   0   0   0   0   0   1   0
#> ENSDARG00000000069   8   5   0   3   2   6   2   4   3   0   0   1   0   1   0
#> ENSDARG00000000086   2   0   1   2   0   2   1   0   0   0   0   1   0   0   0
#>                    X33 X34 X35 X36 X37 X38 X39 X40
#> ENSDARG00000000002   3   1   5   0  12   3   2   1
#> ENSDARG00000000018   1   1   3   3   6   4   3   7
#> ENSDARG00000000019  13  10  21   0  23   3  13   9
#> ENSDARG00000000068   0   0   3   0   0   0   1   0
#> ENSDARG00000000069   2   0   5   0   7   2   5   3
#> ENSDARG00000000086   0   0   1   0   7   2   0   2

When using your own tomo-seq dataset, try to make your data the same structure as the examplary read count matrix. Each row corresponds to a gene and each row correspond to a section. The row names of the matrix are gene names. Importantly, the columns MUST be ordered according to the spatial sequence of sections.

Create an object

Now we create an object representing from the raw read count matrix. Genes expressed in less than 3 sections are filtered out. You can change this threshold by changing the parameter min.section of function createTomo. The output object is an instance of SummarizedExperiment. If you have additional information about sections, save them in colData(object), a data frame used to save meta data describing sections.

zh <- createTomo(zh.data)
#> Normalized count matrix is saved in assay 'normalized'.
#> Scaled count matrix is saved in assay 'scaled'.
zh
#> class: SummarizedExperiment 
#> dim: 12865 40 
#> metadata(0):
#> assays(3): count normalized scaled
#> rownames(12865): ENSDARG00000000002 ENSDARG00000000018 ...
#>   ENSDARG00000095236 ENSDARG00000095580
#> rowData names(1): gene
#> colnames(40): X1 X2 ... X39 X40
#> colData names(1): section

If you have a normalized expression matrix rather than raw read count matrix, it can also be used for input.

your_object <- createTomo(matrix.normalized = normalized) 
# Replace 'normalized' with your normalized expression matrix.

If you have an existing SummarizedExperiment object, createTomo also accepts it as input. Just remember that the object must contain at least one of ‘count’ assay and ‘normalized’ assay.

your_object <- createTomo(se) 
# Replace 'se' with a SummarizedExperiment object.

Normalize and scale data

By default, raw read count matrix is normalized and scaled across sections. The raw read count, normalized read count matrix and scaled read count matrix are saved in ‘count’, ‘normalized’ and ‘scale’ assays of the object. These matrices can be accessed using function assay.

head(assay(zh, 'scaled'), 2)
#>                            X1         X2         X3         X4         X5
#> ENSDARG00000000002  0.2711605 -0.7486253 -0.7486253 -0.7486253 -0.7486253
#> ENSDARG00000000018 -0.7375048 -0.7375048 -0.7375048 -0.7375048 -0.7375048
#>                          X6         X7       X8          X9        X10
#> ENSDARG00000000002 4.725422 -0.7486253 1.145541  0.13187095 0.06301435
#> ENSDARG00000000018 3.300955  0.2058376 1.125715 -0.08792157 2.85520219
#>                          X11        X12        X13       X14        X15
#> ENSDARG00000000002 0.3100379 -0.7486253 -0.7486253 0.7223615  0.2350572
#> ENSDARG00000000018 0.0435205 -0.1800035  0.6251889 0.8903187 -0.3746505
#>                           X16        X17        X18        X19        X20
#> ENSDARG00000000002  1.2367534  0.4586294  0.3047269  0.2002146 -0.7486253
#> ENSDARG00000000018 -0.1097734 -0.4406221 -0.7375048 -0.3875030 -0.1266124
#>                           X21        X22         X23        X24        X25
#> ENSDARG00000000002  1.1313419  0.9039949 -0.52762436 -0.7486253 -0.7486253
#> ENSDARG00000000018 -0.3412363 -0.4327010  0.07770893 -0.3048022 -0.7375048
#>                           X26        X27        X28        X29        X30
#> ENSDARG00000000002 -0.4885558  0.5504044 -0.7486253 -0.7486253 -0.7486253
#> ENSDARG00000000018 -0.3537738 -0.4180532  3.0631010 -0.7375048 -0.7375048
#>                           X31        X32         X33        X34         X35
#> ENSDARG00000000002 -0.7486253  0.3513916 -0.05306309 -0.3883656  0.01307173
#> ENSDARG00000000018 -0.7375048 -0.1964822 -0.56645521 -0.4717243 -0.40034109
#>                           X36        X37         X38        X39        X40
#> ENSDARG00000000002 -0.7486253  0.8666370 -0.10033582 -0.3579938 -0.4763123
#> ENSDARG00000000018  0.2580826 -0.1416776 -0.09980686 -0.3052241  0.6687815

During normalization, the library sizes of all sections are set to the median of all library sizes. They can also be normalized to 1 million counts to obtain Count Per Million (CPM) value by setting parameter normalize.method = "cpm".

zh <- createTomo(zh.data, normalize.method = "cpm") 

We do not normalize gene lengths as we will not perform comparision between two genes. If the normalized read count matrix is used as input, this step is skipped.

Then the normalized data is scaled across sections for each gene. The normalized read counts of each gene are subjected to Z score transformation such that they have mean as 0 and standard deviation as 1.

Correlation analysis

A good start to analyze tomo-seq data is correlation analysis. Here we calculate the Pearson correlation coefficients between every pair of sections across all genes and visualize them with a heatmap. Parameter max.cor defines the maximum value for the heatmap, and coefficients bigger than it are clipped to it. This is because diagonal coefficients are 1, usually much bigger than other coefficients, so clipping them to a smaller value will show other coefficients more clearly.

corHeatmap(zh, max.cor=0.3)

We would expect that adjacent sections have similar transcriptional profiles and thus have bigger correlation coefficients. Therefore, a pair of adjacent sections with small correlation coefficients should be noted. They may act as borders of two zones with different transcriptional profiles. A border of different zones is usually a border of dark blue and light blue/green/yellow on the heatmap. For example, section X13 and X20 are two borders in this dataset according to the heatmap.

Dimensionality reduction analysis

Another method to visualize the similarity of sections is to perform dimensionality reduction. Sections are embedded in a two-dimensional space and plotted as points. similar sections are modeled by nearby points and dissimilar sections are modeled by distant points with high probability.

We first try PCA, a classic linear dimensionality reduction algorithm. We can see a general trend of bottom-left to upper-right with increasing section indexes, but it is hard to find clear borders. The embeddings of sections output by the function are saved in the Tomo object, and you can access them with colData(object).

zh <- runPCA(zh)
#> PC embeddings for sections are saved in column data.
embedPlot(zh, method="PCA")
#> Warning: ggrepel: 6 unlabeled data points (too many overlaps). Consider
#> increasing max.overlaps

head(colData(zh))
#> DataFrame with 6 rows and 3 columns
#>        section       PC1       PC2
#>    <character> <numeric> <numeric>
#> X1          X1 0.0612488 -0.282025
#> X2          X2 0.0524679 -0.274253
#> X3          X3 0.0389170 -0.403186
#> X4          X4 0.0779290 -0.323700
#> X5          X5 0.0510980 -0.407276
#> X6          X6 0.0863790 -0.304588

Next we move to two popular non-linear dimensionality reduction algorithm, tSNE and UMAP. These algorithms are designed to learn the underlying manifold of data and project similar sections together in low-dimensional spaces. Users are welcomed to tune the parameter of these algorithm to show better results with custom dataset.

In the examplary dataset, two clusters of sections with a large margin are shown in both tSNE and UMAP embedding plots. According to the labels of sections, we could identify a border at X21 ~ X22.

set.seed(1)
zh <- runTSNE(zh)
#> TSNE embeddings for sections are saved in column data.
embedPlot(zh, method="TSNE")

zh <- runUMAP(zh)
#> UMAP embeddings for sections are saved in column data.
embedPlot(zh, method="UMAP")
#> Warning: ggrepel: 6 unlabeled data points (too many overlaps). Consider
#> increasing max.overlaps

Clustering analysis

Sometimes it is hard to find borders manually with results above, so we include some clustering algorithms to help users do this.

Hierarchical clustering is good at build a hierachy of clusters. You can easily find similar sections from adjacent nodes in the dendrogram. However, beware that hierarchical clustering is based on greedy algorithm, so its partitions may not be suitable to define a few clusters.

hc_zh <- hierarchClust(zh)
plot(hc_zh)

If certain number of clusters of sections with large margins are observed in embedding plots, or you already decide the number of zones, using K-Means for clustering is a good choice. Input your expected number of clusters as parameter centers, sections will be divided into clusters. The cluster labels output by K-Means are saved in colData(object). When plotting the embeddings of sections, you can use K-Means cluster labels for the colors of sections.

zh <- kmeansClust(zh, centers=3)
#> K-Means clustering labels are saved in colData.
#> between_SS / total_SS =0.83750797944353
head(colData(zh))
#> DataFrame with 6 rows and 8 columns
#>        section       PC1       PC2     TSNE1     TSNE2     UMAP1     UMAP2
#>    <character> <numeric> <numeric> <numeric> <numeric> <numeric> <numeric>
#> X1          X1 0.0612488 -0.282025  -19.1589  -31.9133 -0.465414   5.25085
#> X2          X2 0.0524679 -0.274253  -20.9231  -33.1062 -0.312061   5.37060
#> X3          X3 0.0389170 -0.403186  -18.7993  -35.1827 -0.426064   5.82332
#> X4          X4 0.0779290 -0.323700  -16.7677  -31.9625 -0.721597   5.45968
#> X5          X5 0.0510980 -0.407276  -15.8300  -34.9901 -0.118228   5.53168
#> X6          X6 0.0863790 -0.304588  -13.6882  -30.7051 -1.072831   4.98961
#>    kmeans_cluster
#>         <integer>
#> X1              2
#> X2              2
#> X3              2
#> X4              2
#> X5              2
#> X6              2
embedPlot(zh, group='kmeans_cluster')

Find peak genes

As tomo-seq data contains spatial information, it is important to find spatially expressed genes. These spatially expressed genes may have biological implications in certain zones. We call spatially upregulated genes “peak genes” and a function is used to find these genes. Here are two parameters to judge whether a gene is a peak gene: threshold and length. Genes with scaled read counts bigger than threshold in minimum length consecutive sections are recognized as peak genes.

The output of this function is a data frame containing the names, start section indexes, end section indexes, center section indexes, p values and adjusted p values of peak genes. P values are calculated by approximate permutation tests. Change the parameter nperm to change the number of random permutations.

peak_genes <- findPeakGene(zh, threshold = 1, length = 4, nperm = 1e5)
#> 376peak genes (spatially upregulated genes) are found!
head(peak_genes)
#>                                  gene start end center       p       p.adj
#> ENSDARG00000002131 ENSDARG00000002131     1   4      2 0.01258 0.017454170
#> ENSDARG00000003061 ENSDARG00000003061     1   4      2 0.00188 0.008125057
#> ENSDARG00000003216 ENSDARG00000003216     1   4      2 0.02550 0.027710983
#> ENSDARG00000003570 ENSDARG00000003570     1   4      2 0.00599 0.013817423
#> ENSDARG00000007385 ENSDARG00000007385     1   4      2 0.00188 0.008125057
#> ENSDARG00000008867 ENSDARG00000008867     1   4      2 0.02550 0.027710983

After finding peak genes, we can visualize their expression across sections with a heatmap. Parameter size controls the size of gene names. When there are too many genes and showing overlapping names make the plot untidy, we set it to 0.

expHeatmap(zh, peak_genes$gene, size=0)

Find co-regulated genes

After finding peak genes and taking a look of the output data frame, you may notice that many genes have similar expression pattern. For example, the first 47 peak genes in this dataset all have peak expression at section 1~4. It is intuitive to think that these genes are co-regulated by certain transcription factors and involve in related pathways.

Like what we do for sections, we calculate the Pearson correlation coefficients between every pair of genes across sections and visualize them with a heatmap. Parameter size controls the size of gene names, which is same as that in expHeatmap.

Notice that geneCorHeatmap takes a data frame describing genes as input. You can use the output from findPeakGenes as input for this function. Variables in the data frame can be used to plot a side bar above the heatmap. For example, with default settings, the side bar describe peak centers of genes. Other variables like start can also be used to group genes.

geneCorHeatmap(zh, peak_genes, size=0)

# Use variable 'start' to group genes
geneCorHeatmap(zh, peak_genes, group='start', size=0)

Similarly, we also visualize the two-dimensional embeddings of genes to find clusters of genes with similar expression pattern.

zh <- runTSNE(zh, peak_genes$gene)
#> TSNE embeddings for genes are saved in row data.
geneEmbedPlot(zh, peak_genes)

zh <- runUMAP(zh, peak_genes$gene)
#> UMAP embeddings for genes are saved in row data.
geneEmbedPlot(zh, peak_genes, method="UMAP")

Users can then explore these co-regulated genes to address biological questions.