Contents

1 Installation

if(!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
BiocManager::install("STdeconvolve")

2 Introduction

STdeconvolve is an unsupervised machine learning approach to deconvolve multi-cellular pixel-resolution spatial transcriptomics datasets in order to recover the putative transcriptomic profiles of cell-types and their proportional representation within spatially resolved pixels without reliance on external single-cell transcriptomics references.

2.1 Deconvolution

In this tutorial, we will walk through some of the main functionalities of STdeconvolve.

library(STdeconvolve)

Given a counts matrix from pixel-resolution spatial transcriptomics data where each spatially resolved measurement may represent mixtures from potentially multiple cell-types, STdeconvolve infers the putative transcriptomic profiles of cell-types and their proportional representation within each multi-cellular spatially resolved pixel. Such a pixel-resolution spatial transcriptomics dataset of the mouse olfactory bulb is built in and can be loaded.

data(mOB)
pos <- mOB$pos ## x and y positions of each pixel
cd <- mOB$counts ## matrix of gene counts in each pixel
annot <- mOB$annot ## annotated tissue layers assigned to each pixel

STdeconvolve first feature selects for genes most likely to be relevant for distinguishing between cell-types by looking for highly overdispersed genes across ST pixels. Pixels with too few genes or genes with too few reads can also be removed.

## remove pixels with too few genes
counts <- cleanCounts(counts = cd,
                      min.lib.size = 100,
                      min.reads = 1,
                      min.detected = 1,
                      verbose = TRUE)
## Converting to sparse matrix ...
## Filtering matrix with 262 cells and 15928 genes ...
## Resulting matrix has 260 cells and 14828 genes
## feature select for genes
corpus <- restrictCorpus(counts,
                         removeAbove = 1.0,
                         removeBelow = 0.05,
                         alpha = 0.05,
                         plot = TRUE,
                         verbose = TRUE)
## Removing 124 genes present in 100% or more of pixels...
## 14704 genes remaining...
## Removing 3009 genes present in 5% or less of pixels...
## 11695 genes remaining...
## Restricting to overdispersed genes with alpha = 0.05...
## Calculating variance fit ...
## Using gam with k=5...
## 232 overdispersed genes ...
##  Using top 1000 overdispersed genes.
##  number of top overdispersed genes available: 232

STdeconvolve then applies latent Dirichlet allocation (LDA), a generative statistical model commonly used in natural language processing, to discover K latent cell-types. STdeconvolve fits a range of LDA models to inform the choice of an optimal K.

## Note: the input corpus needs to be an integer count matrix of pixels x genes
ldas <- fitLDA(t(as.matrix(corpus)), Ks = seq(2, 9, by = 1),
               perc.rare.thresh = 0.05,
               plot=TRUE,
               verbose=TRUE)
## Time to fit LDA models was 0.96 mins
## Computing perplexity for each fitted model...
## Time to compute perplexities was 0 mins
## Getting predicted cell-types at low proportions...
## Time to compute cell-types at low proportions was 0 mins
## Plotting...

In this example, we will use the model with the lowest model perplexity.

The shaded region indicates where a fitted model for a given K had an alpha > 1. alpha is an LDA parameter that is solved for during model fitting and corresponds to the shape parameter of a symmetric Dirichlet distribution. In the model, this Dirichlet distribution describes the cell-type proportions in the pixels. A symmetric Dirichlet with alpha > 1 would lead to more uniform cell-type distributions in the pixels and difficulty identifying distinct cell-types. Instead, we want models with alpha < 1, resulting in sparse distributions where only a few cell-types are represented in a given pixel.

The resulting theta matrix can be interpreted as the proportion of each deconvolved cell-type across each spatially resolved pixel. The resulting beta matrix can be interpreted as the putative gene expression profile for each deconvolved cell-type normalized to a library size of 1. This beta matrix can be scaled by a depth factor (ex. 1000) for interpretability.

## select model with minimum perplexity
optLDA <- optimalModel(models = ldas, opt = "min")

## Extract pixel cell-type proportions (theta) and cell-type gene expression
## profiles (beta) for the given dataset.
## We can also remove cell-types from pixels that contribute less than 5% of the
## pixel proportion and scale the deconvolved transcriptional profiles by 1000 
results <- getBetaTheta(optLDA,
                        perc.filt = 0.05,
                        betaScale = 1000)
## Filtering out cell-types in pixels that contribute less than 0.05 of the pixel proportion.
deconProp <- results$theta
deconGexp <- results$beta

2.2 Visualization

We can now visualize the proportion of each deconvolved cell-type across the original spatially resolved pixels.

vizAllTopics(deconProp, pos, 
             groups = annot, 
             group_cols = rainbow(length(levels(annot))),
             r=0.4)
## Plotting scatterpies for 260 pixels with 8 cell-types...this could take a while if the dataset is large.