Reduced dimension

Let us visualize the top 2 markers identified in this way for each cluster.

For reference, the cell type identity associated with markers mentioned in the Guided Clustering Tutorial is indicated in each panel subtitle, where applicable.

library(ggplot2)
library(cowplot)
library(scales)
ggList <- list()
for (markerId in unique(ids(elements(signatures)))) {
    ggDataFrame <- data.frame(
        reducedDim(sce, "TSNE"),
        logcounts=log2(assay(sce, "counts")[markerId, ] + 1)
    )
    markerName <- rowData(sce)[markerId, "Symbol_TENx"]
    seuratIdentity <- subset(as(seuratTutorialMarkers, "data.frame"), element == markerName, "set", drop=TRUE)
    if (length(seuratIdentity) == 0) { seuratIdentity <- " "} # trick to make equal plot dimensions
    gg <- ggplot(ggDataFrame, aes(tSNE_1, tSNE_2, color=logcounts)) +
        geom_point(size=0.1) +
        scale_color_viridis_c() +
        labs(
            title=sprintf("%s (%s)", markerId, markerName),
            subtitle=seuratIdentity,
            x=NULL, y=NULL) +
        theme(
            axis.text = element_blank(), axis.ticks = element_blank(),
            title = element_text(size=rel(0.7)),
            legend.text = element_text(size=rel(0.8)), legend.title = element_text(size=rel(0.8)),
            legend.position = "bottom")
    ggList[[markerId]] <- gg
}
cowplot::plot_grid(plotlist = ggList)

Interactively

It is possible to interactively set the names of each gene signature learned using an interactive Shiny app. For this purpose, we learn as many markers as possible for each cluster, and we pass both the signatures (as a Sets object) and the annotated SingleCellExperiment to the interactive app.

signatures <- learnSignatures(
    sce, assay.type = "counts",
    method = "PositiveProportionDifference", cluster.col="seurat.celltype",
    threshold = 0, n = Inf, min.diff = 0.2)
scePred <- predict(signatures, sce, method="ProportionPositive", cluster.col="seurat.ident")
if (interactive()){
    library(shiny)
    signatures <- runApp(shinyLabels(signatures, scePred))
}

Note how the returned signatures object contains any update you may have done to the names of the gene sets interactively within the app.

Proportion of cells predicted for each signature

One may also visualize the proportion of cells positive for each signature in each cluster. If you renamed any signature in the previous chunk, the predict method should be run again to apply the new labels to the SingleCellExperiment object.

scePred <- predict(signatures, sce, method="ProportionPositive", cluster.col="seurat.ident")
plotProportionPositive(scePred, cluster_columns=FALSE, cluster_rows=FALSE)

Scree detection plot

The learning method "PositiveProportionDifference" allows the trimming of candidate markers to those that are co-detected in a minimal proportion of samples in the target cluster. The trimming process starts from the most frequently detected marker and stop when the frequency of co-detection drops below the required minimal proportion; all candidate markers beyond this point are dropped. The motivation here is that for the purpose of defining qualitative gene signatures, it is useful to ensure that all of the combined gene set that makes up a signature is fully detected in a minimal proportion of samples.

To illustrate this process, the code chunk below computes and displays the proportion of cells in cluster "0" in which the N most frequently detected markers are co-detected. Specifically, the figure displays the proportion of samples with detectable expression of the most frequently detected gene, the two most frequently detected genes (simultaneously), etc.

library(DelayedMatrixStats)
# Choose a cluster for this example
cluster0 <- "0"
topMarkers <- 100
# Fetch only the cells of that cluster
sce0 <- sce[, sce$seurat.ident == cluster0]
# Extract a subset of the most frequently detected genes in those cells
freq0 <- rowSums2(assay(sce0, "counts"))
order0 <- order(freq0, decreasing=TRUE)
# Test whether each marker is detected in each cell
detection0 <- makeMarkerDetectionMatrix(sce0, head(rownames(sce0)[order0], topMarkers))
# Compute the simultaneous detection rate for increasing numbers of markers
scree0 <- makeMarkerProportionScree(detection0)
# Plot
ggplot(scree0, aes(markers, proportion)) +
    geom_line() + geom_point(size=0.4) +
    scale_y_continuous(limits=c(0, 1), labels = scales::percent) +
    scale_x_continuous(limits=c(1, topMarkers)) +
    theme(panel.grid.major=element_line(color="grey90", size=0.5)) +
    labs(x="Number of top markers", y="Proportion positive for all markers")