differential-expression.Rmd

---
title: "Applying CSIDE to Spatial Transcriptomics Data"
author: "Dylan Cable"
date: "December 15th, 2021"
output:
  rmarkdown::html_vignette:
    keep_md: yes
  pdf_document: default
  html_document:
    keep_md: yes
vignette: |
  %\VignetteIndexEntry{differential-expression}
  %\VignetteEncoding{UTF-8}
  %\VignetteEngine{knitr::rmarkdown}
---

```{r setup}

library(spacexr)
library(Matrix)
library(doParallel)
library(ggplot2)
datadir <- system.file("extdata",'SpatialRNA/Vignette',package = 'spacexr') # directory for sample Slide-seq dataset
if(!dir.exists(datadir))
  dir.create(datadir)
savedir <- 'RCTD_results'
if(!dir.exists(savedir))
  dir.create(savedir)
```

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  cache = TRUE,
  out.width = "100%"
)
```

## Introduction

Cell type-Specific Inference of Differential Expression, or CSIDE, is part of the spacexr R package for learning cell type-specific differential expression from spatial transcriptomics data. In this Vignette, we will use CSIDE to test for differential expression in a toy cerebellum Slide-seq dataset. First, we will first use RCTD to assign cell types to a cerebellum Slide-seq dataset. We will define cell type profiles using an annotated single nucleus RNA-sequencing (snRNA-seq) cerebellum dataset. We will test for differential expression across a random explanatory variable.

## Data Preprocessing and running RCTD

First, we run RCTD on the data to annotated cell types. Please note that this follows exactly the content of the spatial transcriptomics RCTD vignette (doublet mode). Please refer to the [spatial transcriptomics vignette](https://raw.githack.com/dmcable/spacexr/master/vignettes/spatial-transcriptomics.html) for more explanation on the RCTD algorithm.

```{r RCTD}
### Load in/preprocess your data, this might vary based on your file type
if(!file.exists(file.path(savedir,'myRCTD.rds'))) {
  counts <- read.csv(file.path(datadir,"MappedDGEForR.csv")) # load in counts matrix
  coords <- read.csv(file.path(datadir,"BeadLocationsForR.csv"))
  rownames(counts) <- counts[,1]; counts[,1] <- NULL # Move first column to rownames
  rownames(coords) <- coords$barcodes; coords$barcodes <- NULL # Move barcodes to rownames
  nUMI <- colSums(counts) # In this case, total counts per pixel is nUMI
  puck <- SpatialRNA(coords, counts, nUMI)
  barcodes <- colnames(puck@counts) # pixels to be used (a list of barcode names). 
  plot_puck_continuous(puck, barcodes, puck@nUMI, ylimit = c(0,round(quantile(puck@nUMI,0.9))), 
                       title ='plot of nUMI') 
  refdir <- system.file("extdata",'Reference/Vignette',package = 'spacexr') # directory for the reference
  counts <- read.csv(file.path(refdir,"dge.csv")) # load in counts matrix
  rownames(counts) <- counts[,1]; counts[,1] <- NULL # Move first column to rownames
  meta_data <- read.csv(file.path(refdir,"meta_data.csv")) # load in meta_data (barcodes, clusters, and nUMI)
  cell_types <- meta_data$cluster; names(cell_types) <- meta_data$barcode # create cell_types named list
  cell_types <- as.factor(cell_types) # convert to factor data type
  nUMI <- meta_data$nUMI; names(nUMI) <- meta_data$barcode # create nUMI named list
  reference <- Reference(counts, cell_types, nUMI)
  myRCTD <- create.RCTD(puck, reference, max_cores = 2)
  myRCTD <- run.RCTD(myRCTD, doublet_mode = 'doublet')
  saveRDS(myRCTD,file.path(savedir,'myRCTD.rds'))
}
```

## Create explanatory variable / covariate

Now that we have successfully run RCTD, we can create a explanatory variable (i.e. covariate) used for predicting differential expression in CSIDE. In this case, the variable, `explanatory.variable`, will be randomly generated, but in general one should set the explanatory variable to biologically relevant predictors of gene expression such as spatial position.
The explanatory variable itself is a vector of values, constrained between 0 and 1, with names matching the pixel names of the `myRCTD` object.

Here, we also artifically upregulate the expression of the Lsamp gene (in regions of high explanatory variable) to see whether CSIDE can detect this differentially expressed gene.

```{r SpatialRNA, results = 'hide', fig.height = 6, fig.width = 6}
### Create SpatialRNA object
myRCTD <- readRDS(file.path(savedir,'myRCTD.rds'))
set.seed(12345)
explanatory.variable <- runif(length(myRCTD@spatialRNA@nUMI))
names(explanatory.variable) <- names(myRCTD@spatialRNA@nUMI) # currently random explanatory variable 
print(head(explanatory.variable))

#Differentially upregulate one gene
change_gene <- 'Lsamp'
high_barc <- names(explanatory.variable[explanatory.variable > 0.5])
low_barc <- names(explanatory.variable[explanatory.variable < 0.5])
myRCTD@originalSpatialRNA@counts[change_gene, high_barc] <- myRCTD@spatialRNA@counts[change_gene, high_barc] * 3 

plot_puck_continuous(myRCTD@spatialRNA, names(explanatory.variable), explanatory.variable, ylimit = c(0,1), title ='plot of explanatory variable') 
```

## Running CSIDE

After creating the explanatory variable, we are now ready to run CSIDE using the `run.CSIDE.single` function. We will use two cores, and a false discovery rate of 0.25. Next, we will set a gene threshold (i.e. minimum gene expression) of 0.01, and we will set a cell_type_threshold (minimum instances per cell type) of 3.

**Warning**: On this toy dataset, we have made several choices of parameters that are not recommended for regular use. On real datasets, we recommend first consulting the CSIDE default parameters. This includes `gene_threshold` (default 5e-5), `cell_type_threshold` (default 125), and `fdr` (default 0.01). Please see `?run.CSIDE.single` for more information on these parameters. 

```{r DEgenes}
#de
myRCTD@config$max_cores <- 2
myRCTD <- run.CSIDE.single(myRCTD, explanatory.variable, gene_threshold = .01, 
                        cell_type_threshold = 3, fdr = 0.25) 

saveRDS(myRCTD,file.path(savedir,'myRCTDde.rds'))
```

## CSIDE results

After running CSIDE, it is time to examine CSIDE results. For each cell type, a results dataframe for significant genes is stored in `myRCTD@de_results$sig_gene_list`. In particular, notice the columns `Z_score` (Z-score), `log_fc` (estimated DE loge-fold-change), and `p_val` (p-value). We also have the mean and standard errors of loge expression in each of the two regions (`mean_0`, `mean_1`, `sd_0`, and `sd_1`). We will focus on cell type 10. Furthermore, we will examine the original Lsamp gene, which was detected to be significantly differentially expressed in cell type 10. CSIDE model fits for all genes are stored in `myRCTD@de_results$gene_fits`, and we demonstrate below how to access the point estimates, standard errors, and convergence.

```{r CSIDE_results,  fig.width = 8, fig.height=8}
#print results for cell type '18'
cell_type <- '10'
results_de <- myRCTD@de_results$sig_gene_list[[cell_type]]
print(results_de)
sig_gene <- change_gene
print(paste("following results hold for", sig_gene))
print("check for covergence of each cell type")
print(myRCTD@de_results$gene_fits$con_mat[sig_gene, ]) 
print('estimated DE')
print(myRCTD@de_results$gene_fits$mean_val[sig_gene, ]) 
print('standard errors for non-intercept terms')
print(myRCTD@de_results$gene_fits$s_mat[sig_gene, ]) 
```

Finally, we will plot CSIDE results in the `savedir` directory!

The following plot shows a spatial visualization of the Lsamp gene, which was determined to be differentially expressed.

The function `make_all_de_plots` will automatically generate several types of plots displaying the CSIDE results across all genes and cell types.

```{r results,  fig.width = 8, fig.height=8}
myRCTD <- readRDS(file.path(savedir,'myRCTDde.rds'))
plot_gene_two_regions(myRCTD, sig_gene, cell_type, min_UMI = 10)
make_all_de_plots(myRCTD, savedir)
```