, "dgTMatrix")' is deprecated.
## Use 'as(., "TsparseMatrix")' instead.
## See help("Deprecated") and help("Matrix-deprecated").
```
```
## - Filtering features:
```
```
## - Selected features: 12514
```
```
## - Discarded features: 1434
```
```
##
```
```
## === Processing single-cell data
```
```
## - Removing 16 genes without expression in any cell
```
```
## - Filtering features:
```
```
## - Selected features: 12350
```
```
## - Discarded features: 488
```
```
##
## === Number of shared genes between single-cell and spatial transcriptomics datasets: 10941
```
```
## - Original # genes in single-cell data: 12350
```
```
## - Original # genes in ST data (object with the greatest # genes): 12514
```
```
##
## === Number of removed mitochondrial genes: 11
```
```
##
## === Number of genes after filtering based on logFC: 1041
```
```
##
## === Final number of dimensions for further analyses: 1041
```
We can show some basic information about the object:
```r
mouseDLN.SDDLS
```
```
## An object of class SpatialDDLS
## Real single-cell profiles:
## 1041 features and 14989 cells
## rownames: Kdr Acsbg1 Cx3cr1 ... Cx3cr1 Cd14 H2afj AA467197
## colnames: AAGCCATCAGGAGGTT-1-1 TTCCACGCAAACACGG-1-2 GAGATGGAGGTTATAG-1-0 ... GAGATGGAGGTTATAG-1-0 GGCGTCAGTTGGTAGG-1-3 TGTGAGTGTGTCCAAT-1-2 TTCGATTGTAAGGAGA-1-0
## Spatial experiments:
## 1 experiments
## 1041 features and 1092 spots
## rownames: Klri2 Mmp25 Cacna1e ... Cacna1e F630028O10Rik Map4k4 Clec4a1
## colnames: CAATAAACCTTGGCCC-1-1 AACAGGAAATCGAATA-1-1 CTGGCGGGAATAAGTA-1-0 ... CTGGCGGGAATAAGTA-1-0 TTCCTCTGCCCGAATA-1-0 CTATCGGGTCTCAACA-1-1 ATCCTGCGTGGAATGG-1-1
## Project: SpatialDDLS-Proj
```
In this case, we are only working on 1 spatial transcriptomics dataset, but an arbitrary number of `SpatialExperiment` objects can be loaded.
## Simulation of mixed transcriptional profiles
Now, we are going to simulate cell composition matrices that will serve to simulate mixed transcriptional profiles with known cell proportions. This is done by the `genMixedCellProp` function in which we can control different aspects, such as the number of mixed transcriptional profiles to be generated or the number of cells used to simulate each mixed profile. These parameters must be decided depending on the size of the single-cell reference and the available computational resources. For this example, and as standard reference, we will use `num.sim.spots = 10000` and `n.cells = 50`. The cell type composition of these mixed profiles will be generated by three methods:
* A random sampling of a Dirichlet distribution. Within this set of samples, in order to make these proportions more sparse, the `prob.sparity` parameter controls the probability of having missing cell types in each simulated spot, as opposed to a mixture of all cell types.
* Pure mixed transcriptional profiles composed of `n.cells` cells of the same cell type aggregated.
* Transcriptional profiles in which a minimum number of missing cell types will be imposed. This is controlled by the `min.zero.prop` argument.
The relative abundance of samples generated by these criteria can be controlled by the `proportion.method` parameter. Finally, the `genMixedCellProp` function will automatically divide the reference cell profiles contained in the `single.cell.real` slot into training and test subsets and randomly assign `n.cells` cells to generate every mixed transcriptional profile.
```r
mouseDLN.SDDLS <- genMixedCellProp(
mouseDLN.SDDLS,
cell.ID.column = "CellID",
cell.type.column = "broad_cell_types",
num.sim.spots = 10000,
n.cells = 50,
min.zero.prop = 5,
balanced.type.cells = TRUE
)
```
```
##
## === The number of mixed profiles that will be generated is equal to 10000
```
```
##
## === Training set cells by type:
```
```
## - B cells: 6269
## - CD4 T cells: 1532
## - CD8 T cells: 2451
## - cDC1s: 75
## - cDC2s: 65
## - GD T cells: 66
## - Macrophages: 79
## - Migratory DCs: 224
## - Monocytes: 59
## - NK cells: 70
## - pDCs: 58
## - Tregs: 292
```
```
## === Test set cells by type:
```
```
## - B cells: 2090
## - CD4 T cells: 511
## - CD8 T cells: 817
## - cDC1s: 25
## - cDC2s: 22
## - GD T cells: 22
## - Macrophages: 26
## - Migratory DCs: 75
## - Monocytes: 20
## - NK cells: 23
## - pDCs: 20
## - Tregs: 98
```
```
## === Probability matrix for training data:
```
```
## - Mixed spots: 7500
## - Cell types: 12
```
```
## === Probability matrix for test data:
```
```
## - Mixed spots: 2500
## - Cell types: 12
```
```
## DONE
```
Then, we can call the `simMixedProfiles` function, which will generate the actual mixed transcriptional profiles using the cell composition matrices generated in the previous step. This step may take a while depending on the number of transcriptional profiles to be simulated, so be patient! In addition, users can choose the method by which the mixed profiles will be generated. We recommend summing up raw counts, and then normalizing samples to obtain logCPMs (`mixing.function = "AddRawCount"`), but other methods are available (see Documentation).
```r
mouseDLN.SDDLS <- simMixedProfiles(mouseDLN.SDDLS, threads = 3)
```
```
## === Setting parallel environment to 3 thread(s)
```
```
##
## === Generating train mixed profiles:
```
```
##
## === Generating test mixed profiles:
```
```
##
## DONE
```
## Training a fully-connected neural network using mixed transcriptional profiles
After generating a set of mixed transcriptional profiles with known cell composition, we can then train a neural network using the training subset and evaluate the model by predicting cell type proportions on the test subset. The trained model can deconvolute the cellular composition of new transcriptional profiles, such as spots in a spatial transcriptomics experiment. The architecture of the network is fully customizable, although in our experience, the default hyperparameters used in this example work for most of the cases. Particularly, we will employ a model with two hidden layers, each consisting of 200 neurons, and a training process involving 60 epochs.
```r
mouseDLN.SDDLS <- trainDeconvModel(
mouseDLN.SDDLS,
verbose = FALSE
)
```
```
##
1/79 [..............................] - ETA: 9s - loss: 0.0840 - accuracy: 0.8750 - mean_absolute_error: 0.0175 - categorical_accuracy: 0.8750
79/79 [==============================] - 0s 622us/step - loss: 0.0818 - accuracy: 0.8712 - mean_absolute_error: 0.0170 - categorical_accuracy: 0.8712
##
79/79 [==============================] - 0s 629us/step - loss: 0.0818 - accuracy: 0.8712 - mean_absolute_error: 0.0170 - categorical_accuracy: 0.8712
```
Some metrics about the training progress can be shown by setting `verbose = TRUE` or by calling the object:
```r
mouseDLN.SDDLS
```
```
## An object of class SpatialDDLS
## Real single-cell profiles:
## 1041 features and 14989 cells
## rownames: Clec4a1 Csf1r Sapcd2 ... Sapcd2 Cd79a Fam20c Rpl37a
## colnames: AGGGTTTTCGTAACCA-1-1 ACGGGTCGTTTACGAC-1-2 GCCATTCGTAGGCAGT-1-2 ... GCCATTCGTAGGCAGT-1-2 TTACAGGAGGAGAGTA-1-3 GGGTCACCACAACGAG-1-0 GACTCAACATGACACT-1-2
## Spatial experiments:
## 1 experiments
## 1041 features and 1092 spots
## rownames: Klrc1 Rps3 Ankrd33b ... Ankrd33b Iglc3 Plekha4 Lynx1
## colnames: CAATGGATCTCTACCA-1-1 GGCGCGTTCGAGTTTA-1-1 GAAGCGTGAGGAATTT-1-1 ... GAAGCGTGAGGAATTT-1-1 GTGGAACCTACATGCG-1-1 AACTTGCCCGTATGCA-1-0 ATAAAGGCTCGGTCGT-1-1
## Cell type composition matrices:
## Cell type matrix for traindata: 7500 bulk samples and 12 cell types
## Cell type matrix for testdata: 2500 bulk samples and 12 cell types
## Simulated mixed spots:
## train spots:
## 1041 features and 7500 spots
## rownames: Ngfr Trib1 Timp2 ... Timp2 Rps3a1 Pou2af1 H3f3a
## colnames: Spot_train_7037 Spot_train_7151 Spot_train_5337 ... Spot_train_5337 Spot_train_3547 Spot_train_2476 Spot_train_1594
## test spots:
## 1041 features and 2500 spots
## rownames: Eif5a Spint1 Ip6k3 ... Ip6k3 Cox6a2 Nudt17 Gsr
## colnames: Spot_test_2242 Spot_test_2302 Spot_test_1868 ... Spot_test_1868 Spot_test_1775 Spot_test_1287 Spot_test_385
## Trained model: 60 epochs
## Training metrics (last epoch):
## loss: 0.0855
## accuracy: 0.8157
## mean_absolute_error: 0.0193
## categorical_accuracy: 0.8157
## Evaluation metrics on test data:
## loss: 0.0818
## accuracy: 0.8712
## mean_absolute_error: 0.017
## categorical_accuracy: 0.8712
## Project: SpatialDDLS-Proj
```
Anyhow, more advanced metrics can be calculated using the `calculateEvalMetrics` function. This function computes mean absolute error (MAE) and mean squared error (MSE) metrics per cell type, providing insight into the model's performance for each cell type. These metrics can be visualized using various functions:
```r
mouseDLN.SDDLS <- calculateEvalMetrics(mouseDLN.SDDLS)
distErrorPlot(
mouseDLN.SDDLS,
error = "AbsErr",
x.by = "CellType",
color.by = "CellType",
error.labels = FALSE,
type = "boxplot",
size.point = 0.5
)
```
```r
distErrorPlot(
mouseDLN.SDDLS,
x.by = "pBin",
error = "AbsErr",
facet.by = "CellType",
color.by = "CellType",
error.label = TRUE,
type = "boxplot"
)
```
```r
corrExpPredPlot(
mouseDLN.SDDLS,
color.by = "CellType",
facet.by = "CellType",
corr = "both",
size.point = 0.5
)
```
```
## `geom_smooth()` using formula = 'y ~ x'
```
As it can be seen, the overall performance is satisfactory, indicating that the model is capable of identifying the distinctive features of each cell type to provide precise predictions of the cell type composition of transcriptional profiles.
## Deconvolution of the spatial transcriptomics dataset
Finally, we can use our trained model to deconvolute the signals of each spot using the `deconvSpatialDDLS` function. By default, this function uses the trained model to predict cell proportions of two sets of transcriptional profiles obtained from the ST datasets:
* 'Intrinsic' profiles: these are the actual transcriptional profiles of every spot in the ST dataset.
* 'Extrinsic' profiles: these are simulated profiles generated from the surrounding spots of every spot. The concept is to create a set of transcriptional profiles that represent the transcriptional features of the spatial context of each spot.
The latter can be used to understand how similar each spot is to its neighbors. Considering the hypothesis that we can infer the cellular composition of each spot based on its surroundings given the correlation between spatial location and cell composition/transcriptional features, we can use this information to spatially contextualize our predictions and improve their accuracy. We refer to this process as spatial regularization. Details about the methodology are explained in the Documentation and @Mananes2023.
```r
mouseDLN.SDDLS <- deconvSpatialDDLS(
mouseDLN.SDDLS, index.st = 1, k.spots = 6, fast.pca = TRUE
)
```
```
## === Normalizing data (LogCPM)
```
```
## === Predicting cell type proportions
```
```
##
1/35 [..............................] - ETA: 0s
35/35 [==============================] - 0s 551us/step
##
35/35 [==============================] - 0s 563us/step
```
```
##
## === Calculating distances in PCA space
```
```
##
## === Calculating 50 PCs
```
```
## === Calculating alpha factors based on distances
```
```
## DONE
```
Now, let's project these predicted proportions in the spatial coordinates:
```r
plotSpatialPropAll(mouseDLN.SDDLS, index.st = 1)
```
To reveal hidden patterns in the coordinates caused by using the same color scale, we can utilize the `plotSpatialProp` function to independently plot each cell type:
```r
list.plots <- lapply(
X = trained.model(mouseDLN.SDDLS) %>% cell.types(),
FUN = \(x) {
plotSpatialProp(
mouseDLN.SDDLS, index.st = 1, cell.type = x, size.point = 1,
colors = "blues"
) + coord_fixed()
}
)
ggarrange(plotlist = list.plots[1:4], align = "hv")
```
```r
ggarrange(plotlist = list.plots[5:8], align = "hv")
```
```r
ggarrange(plotlist = list.plots[9:12], align = "hv")
```
In addition to the 'regularized' cell proportions, we can plot the predictions calculated for the intrinsic and extrinsic transcriptional profiles. For instance, let's plot those predicted from the extrinsic transcriptional profiles:
```r
list.plots <- lapply(
X = trained.model(mouseDLN.SDDLS) %>% cell.types(),
FUN = \(x) {
plotSpatialProp(
mouseDLN.SDDLS, index.st = 1, cell.type = x, size.point = 1,
colors = "blues", prediction = "Extrinsic"
) + coord_fixed()
}
)
ggarrange(plotlist = list.plots[1:4], align = "hv")
```
```r
ggarrange(plotlist = list.plots[5:8], align = "hv")
```
```r
ggarrange(plotlist = list.plots[9:12], align = "hv")
```
As one may expect, the extrinsic predictions are a smoothed version of the final ones. It is also possible to visualize distances between the extrinsic and intrinsic transcriptional profiles of each spot to understand how the regularization step works by using the `plotDistances` function:
```r
plotDistances(mouseDLN.SDDLS, index.st = 1, size.point = 1.5) + coord_fixed()
```
Those spots with distances less than the mean distance were regularized according to their nearest neighbor spots.
## Interpreting the model
In order to make predictions more transparent, `SpatialDDLS` includes an additional module designed to provide insights into the model's decision-making process. It relies on calculating the predicted classes/loss function gradients with respect to the input variables, a method popularly known as Vanilla Gradient. These numeric values are computed for each gene and cell type using the pure mixed transcriptional profiles previously simulated. Therefore, they can be interpreted as the extent to which each feature is contributing to the model's predictions. While these values are initially calculated at the sample/gene level, they are aggregated at the cell type level in order to assess the relevance of each gene to each cell type prediction. These steps are performed through the `interGradientDL` function:
```r
mouseDLN.SDDLS <- interGradientsDL(
mouseDLN.SDDLS, method = "class", scaling = "standardize"
)
```
Importantly, depending on the `method` parameter, positive and negative gradients must be differently interpreted:
* If gradients with respect to the input variables were calculated using the loss function (`method = "loss"`), genes with negative gradients (those that minimize the loss function) will be positively correlated with the presence of each cell type.
* Conversely, if gradients with respect to the input variables were calculated using classes (`method = "class"`), genes with positive gradients (those that make the probability of being a cell type higher) will be positively associated with each cell type.
It is important to note that these markers should not be interpreted as cell type markers. Rather, they serve as indications to help interpret the model's performance. In addition, due to the multivariate nature of our approach, gradients are surrogates at the feature level for predictions made considering all input variables collectively, and thus caution should be exercised in drawing direct conclusions about specific gene-cell type relationships.
For this example, let's calculate gradients of the class function with respect to the input features and show the top 5 genes with the greatest gradients per cell type:
```r
top.gradients <- topGradientsCellType(
mouseDLN.SDDLS, method = "class", top.n.genes = 5
)
sapply(
top.gradients, \(x) x$Positive
) %>% as.data.frame()
```
```
## B cells CD4 T cells CD8 T cells cDC1s cDC2s GD T cells Macrophages Migratory DCs Monocytes NK cells pDCs Tregs
## 1 Dmxl1 Igfbp4 Cd8b1 Slc16a9 Ddx5 Gm42418 Svbp Sec11c Kif22 Runx3 Kif22 Foxp3
## 2 Tnfrsf8 Satb1 Acp5 Tnfrsf8 Rpl4 Nrgn Fth1 Asprv1 Samhd1 Ccr9 Rpl31 Stk39
## 3 Cd79b Ly6c1 Ccr9 1500009L16Rik Samhd1 Ddx5 Chchd10 Ctsw Ndfip1 Mbnl1 Iglv1 Ctla4
## 4 Npm1 Lef1 Rgcc Cox7b Lefty1 Kif22 Atp6v1g1 Kif22 Gpr83 Cd72 Emp3 Itih5
## 5 Fam20c Rps29 Runx3 Cdc20 Gm26917 Il2rg Pla2g2d Cd164 Ddx5 Il2rg Rac2 Fyn
```
As can be seen, among the top 5 genes some canonical markers for different cell types appear, such as Cd8 for CD8 T cells or Foxp3 for Tregs. These are just the top 5 genes, so considering a higher number of genes can provide a more comprehensive understanding of the genes being used by the model.
We also provide the `plotHeatmapGradsAgg` function for visualizing the top N mean gradients per cell type. This plot highlights genes with high gradients across different cell types, reflecting the multivariate nature of neural networks.
```r
hh <- plotHeatmapGradsAgg(mouseDLN.SDDLS, top.n.genes = 4, method = "class")
hh$Absolute
```
Finally, we can use the `plotSpatialGeneExpr` function to visualize the spatial distribution of the top N genes per cell type in the ST dataset. Let's plot some genes for some cell types just for demonstration purposes:
```r
top.genes <- topGradientsCellType(mouseDLN.SDDLS, top.n.genes = 4)
for (i in c("B cells", "CD4 T cells", "CD8 T cells", "Tregs", "Monocytes")) {
list.plots <- list()
for (j in top.genes[[i]][["Positive"]]) {
list.plots[[j]] <- plotSpatialGeneExpr(
mouseDLN.SDDLS, index.st = 1, gene = j, size.point = 0.5,
title = paste0(i, " - ", j)
) + coord_fixed() + theme(legend.position = "none") ## legend removed just for viz
}
print(ggpubr::ggarrange(plotlist = list.plots, align = "hv"))
}
```
## Clustering analysis
The `SpatialDDLS` R package also includes some functions to cluster the ST dataset according to the predicted cell composition of each spot. This functionality enables to dissect the ST datasets into distinct cellular niches, information that might be relevant for further analyses.
```r
mouseDLN.SDDLS <- spatialPropClustering(mouseDLN.SDDLS, k.nn = 20)
```
```
## No 'index.st' provided. Deconvoluting all SpatialExperiment objects contained in the `spatial.experiments` slot
```
```
## === Selected graph-based clustering
```
```
## === Running clustering for slide 1
```
```r
plotSpatialClustering(mouseDLN.SDDLS) + coord_fixed()
```
```
## No 'index.st' provided. Using first ST dataset
```
```
## === Plotting first clustering configuration Clustering.graph.k.20
```
## Comparing deconvoluted cell proportions with colocalization of cell markers
Finally, we are going to assess whether there is a collocation between the predicted cell type proportions and the expression of known markers for each cell type. This analysis aims to validate the model's predictions by comparing them with well-established cellular markers, but it does not mean to be a quantitative validation of the model.
```r
customMarkers <- list(
"B cells" = c("Cd74", "Cd19", "Cd79a", "Cd79b", "Ly6d"),
"CD4 T cells" = c("Cd4", "Lef1", "Fyb"),
"CD8 T cells" = c("Cd8b1", "Cd8a", "Trac"),
cDC1s = c("Xcr1", "Irf8"),
cDC2s = c("Irf4", "Cd4"),
"GD T cells" = c("Il7r", "Id2"),
Macrophages = c("Lyz2", "Lyz1", "Cd86", "Ly6c1"),
"Migratory DCs" = c("Ccl5", "Anxa3", "Fscn1"),
Monocytes = c("Fcer1g", "Cst3", "Lst1", "Itgam", "Kit", "Fcgr3"),
"NK cells" = c("Nkg7", "Il2rb", "Gzma"),
pDCs = c("Siglech", "Plac8", "Ly6c2", "Vtsb", "Zeb2", "Siglech"),
Tregs = c("Ikzf2", "Il2ra", "Foxp3")
) %>% lapply(FUN = function(x) x[x %in% rownames(MouseDLN.ST)])
```
```r
## calculate z-scores
exprST <- MouseDLN.ST@assays@data[[1]]
logCPM <- edgeR::cpm(exprST, log = TRUE)
meanZscoresCustom <- purrr::map(
.x = names(customMarkers),
.f = ~{ colMeans(t(scale(t(logCPM[customMarkers[[.x]], , drop = FALSE])))) }
) %>% do.call(cbind, .)
colnames(meanZscoresCustom) <- names(customMarkers)
```
```r
color.z.scores <- rev(
colorRampPalette(RColorBrewer::brewer.pal(n = 10, name = "RdBu"))(20)
)
st.coor <- SpatialExperiment::spatialCoords(
spatial.experiments(object = mouseDLN.SDDLS, index.st = 1)
)
colnames(st.coor) <- paste("Spatial", 1:2)
dfPlotLong <- reshape2::melt(
as.data.frame(cbind(st.coor, meanZscoresCustom)),
id.vars = c("Spatial 1", "Spatial 2"),
variable.name = "CellType", value.name = "Zscore"
)
dfPlotLong %>% ggplot(
aes(x = .data[["Spatial 1"]], y = .data[["Spatial 2"]], color = Zscore)
) + geom_point(size = 0.5) + theme_classic() +
ggtitle("Mean z-score of cell type markers") +
scale_color_gradientn(colors = color.z.scores, limit = c(-2, 2)) +
theme(
plot.title = element_text(face = "bold", hjust = 0.5),
axis.title.x = element_blank(), axis.text.x = element_blank(),
axis.ticks.x = element_blank(), axis.title.y = element_blank(),
axis.text.y = element_blank(), axis.ticks.y = element_blank()
) + coord_fixed() + facet_wrap(~ CellType)
```
As it can be seen, markers for each cell type colocalize with the cell proportions predicted by `SpatialDDLS`, demonstrating its ability to deconvolute ST samples. For more examples and a quantitative assessment of the algorithm, please see the published manuscript @Mananes2023.
## References
