mining_pd_genetics_imaging_3.Rmd

---
title: "Mining genetic, transcriptomic and imaging data in Parkinson’s disease - 3"
author: "Manuel Tognon"
date: "8/10/2021"
output:
html_document: default
---

# Functional interpretation

Differential expression analysis on putamen and caudate dopamine uptake values.
The below differential expression analysis test for genes showing a 
differential expression by DaTscan putamen and caudate values, showed by the 
individuals participating to PPMI baseline RNA-seq.
After the differential expression analysis we'll be carried out an enrichment
analysis to recover the  biological pathways potentially perturbed by 
differentially expressed genes.

R packages required:

   - DESeq2
   - ggplot2
   - limma
   - fgsea
   - org.Hs.eg.db
   - ReactomePA
   - data.table
   - RNOmni


```{r}

if(!require("BiocManager", character.only = TRUE))
{
  install.packages("BiocManager")
  if(!require("BiocManager", character.only = TRUE))
  {
    stop("BiocManager package not found")
  }
}

if(!require("DESeq2", character.only = TRUE))
{
  BiocManager::install("DESeq2")
  if(!require("DESeq2", character.only = TRUE))
  {
    stop("DESeq2 package not found")
  }
}

if(!require("ggplot2", character.only = TRUE))
{
  install.packages("ggplot2")
  if(!require("ggplot2", character.only = TRUE))
  {
    stop("ggplot2 package not found")
  }
}

if(!require("limma", character.only = TRUE))
{
  BiocManager::install("limma")
  if(!require("limma", character.only = TRUE))
  {
    stop("limma package not found")
  }
}

if(!require("fgsea", character.only = TRUE))
{
  BiocManager::install("fgsea")
  if(!require("fgsea", character.only = TRUE))
  {
    stop("fgsea package not found")
  }
}

if(!require("org.Hs.eg.db", character.only = TRUE))
{
  BiocManager::install("org.Hs.eg.db")
  if(!require("org.Hs.eg.db", character.only = TRUE))
  {
    stop("org.Hs.eg.db")
  }
}

if(!require("ReactomePA", character.only = TRUE))
{
  BiocManager::install("ReactomePA")
  if(!require("ReactomePA", character.only = TRUE))
  {
    stop("ReactomePA package not found")
  }
}

if(!require("data.table", character.only = TRUE))
{
  install.packages("data.table")
  if(!require("data.table", character.only = TRUE))
  {
    stop("data.table package not found")
  }
}

if(!require("RNOmni", character.only = TRUE))
{
  install.packages("RNOmni")
  if(!require("RNOmni", character.only = TRUE))
  {
    stop("RNOmni package not found")
  }
}

if(!require("STRINGdb", character.only = TRUE))
{
  install.packages("STRINGdb")
  if(!require("STRINGdb", character.only = TRUE))
  {
    stop("STRINGdb package not found")
  }
}

```


Import the libraries previously installed.

```{r}
library(DESeq2)
library(ggplot2)
library(limma)
library(fgsea)
library(org.Hs.eg.db)
library(ReactomePA)
library(data.table)
library(RNOmni)
library(STRINGdb)
library(httr)
```



Let's begin by importing data RNA-seq counts and phenotype data!

```{r}

counts <- read.table("data/transcriptomic/PPMI-RNA-seq_baseline_rawCounts.txt")
metadata <- read.csv("data/pheno/PPMI-baseline_pheno.csv")
```


Samples are identified by complex names, let's recover their PATNO...

```{r}
getid <- function(s){
  patno <- unlist(strsplit(s, ".", fixed = T))[4]
  return(patno)
}
samples.ids <- colnames(counts)
samples.ids.new <- sapply(samples.ids, getid)
names(samples.ids.new) <- NULL  # remove old names
colnames(counts) <- samples.ids.new
head(counts)
```
```{r}
# sort the samples in metadata 
rownames(metadata) <- metadata$PATNO
metadata <- metadata[colnames(counts),]
head(metadata)
```

Before proceeding with DE analysis, let's cast our covariates to factors:

```{r}
metadata$gen <- as.factor(metadata$gen)
metadata$age_cat <- as.factor(metadata$age_cat)
metadata$educ <- as.factor(metadata$educ)
```


Then, we categorize in two groups normalized Putamen Left values.

```{r}
metadata$PUTAMEN_L.cat <- cut(
  metadata$PUTAMEN_L_norm, 
  breaks = c(quantile(metadata$PUTAMEN_L_norm, probs = seq(0, 1, by = 1/2))),
  labels = c(1,2)
)
metadata$PUTAMEN_L.cat[34] <- 1  # fix quantile() bug

```


Now we are ready to create the statistical model and the DESeq object for DE analysis!

```{r}
dds <- DESeqDataSetFromMatrix(
  countData = counts, colData = metadata, 
  design = ~ PUTAMEN_L.cat + gen + age_cat + ENROLL_CAT + educ
)
```


Low quality genes are filtered out keeping only those expressed in at least one sample for each group.

```{r}
summary(metadata$PUTAMEN_L.cat)  # 130  130
keep <- apply(counts(dds), 1, function(x){sum(x != 0) > 130})
dds <- dds[keep,]
```

First, we now check for potential batch effects or outliers in our data

```{r}
# count normalization 
vsd <- vst(dds, blind = F)

# plot PCA coloring samples by gender group
p <- plotPCA(vsd, intgroup = "gen", returnData=F)
p + geom_label(aes(label = colnames(dds))) + 
  ggtitle("Gender")  #  batch effect by gender
```

There is a serious batch effect caused by samples geneder...
We should remove it before proceeding, since it could seriously bias our results!

````{r}
# correct counts to remove batch effect
assay(vsd) <- limma::removeBatchEffect(assay(vsd), vsd$gen)
p <- plotPCA(vsd, intgroup = "gen", returnData=F)
p + geom_label(aes(label = colnames(dds))) + 
  ggtitle("Gender")  # batch effect removed
```


Once the batch effect has been removes, we can perform differential expression analysis

```{r}
dds <- DESeq(dds)
```

To visualize the results we can use the results() function of the "DESeq2" package specifying the appropriate contrast.
Then, we remove non significant genes (adjusted p-value < 0.1) and we order them by LogFC for a better visualization.

```{r}
results.putamen.l <- results(dds, contrast = c("PUTAMEN_L.cat", "1", "2"))
results.putamen.l <- results.putamen.l[!is.na(results.putamen.l$padj),]
results.putamen.l <- results.putamen.l[results.putamen.l$padj < .1,]
dim(results.putamen.l)
results.putamen.l <- results.putamen.l[order(results.putamen.l$log2FoldChange, decreasing = T),]

head(results.putamen.l)
```


Once that the DE genes were retrieved, we are ready to perform Enrichment analysis using the "fgsea" package.
As first step we map Ensembl IDs of genes to the EntrezID as required by "fgsea".

```{r}
# remove the ID versions from the genes' names
ensembl.ids <- rownames(results.putamen.l)
ensembl.ids <- sapply(
  rownames(results.putamen.l),
  function(x){
    unlist(strsplit(x, fixed = T, split = "."))[1]
  }
)
names(ensembl.ids) <- NULL  # remove old names 

# mapping 
entrez.ids <- select(
  org.Hs.eg.db, 
  keys = ensembl.ids, 
  columns = "ENTREZID",
  keytype = "ENSEMBL"
)
entrez.ids <- entrez.ids[!duplicated(entrez.ids$ENSEMBL),]  # remove duplicates
entrez.ids <- entrez.ids[!is.na(entrez.ids$ENTREZID),]  # remove NAs
results.putamen.l <- results.putamen.l[entrez.ids$ENSEMBL,]  # remove unmapped genes

# change the names of the genes
rownames(results.putamen.l) <- entrez.ids$ENTREZID
head(results.putamen.l)
```


Now, we can perform Enrichment analysis.
First, we order genes by LogFC. 
Then, we retrieve the Reactome gene sets using the ```reactomePathways()``` function and then we run ```fgsea()```.
At the end we filter out non significant pathways (adjusted p-value < 0.1)

```{r}
# construct a vector containing the LogFC of each gene
lfc <- results.putamen.l$log2FoldChange
names(lfc) <- rownames(results.putamen.l)

# recover pathways 
pathways <- reactomePathways(names(lfc))
enr.fgsea.reactome <- fgsea(
  pathways, 
  lfc,
  nPermSimple = 1000000,
  eps = 0,
  minSize = 5
)

# subset to signifcant pathways (q-value < 0.1)
enr.fgsea.reactome <- enr.fgsea.reactome[enr.fgsea.reactome$padj < .1,]
head(as.data.frame(enr.fgsea.reactome))
```

To export the pathways retrieved we can use the fwrite() function of the package "data.table"

```{r, eval=FALSE}
# store results
fwrite(
  as.data.frame(enr.fgsea.reactome),
  "../resultsRNAseq/Fgsea_PutamenL_Reactome.csv",
  sep = "\t",
  sep2 = c("", " ", "")
)
```

We can use the same code for all the features we considered in our study.
Therefore, we will get a CSV file with potentially perturbed pathways for each neuroimaging feature we 
considered.

And now?

Let's combine the results obtained from GWAS, DE and Enrichment analyses! 

In our example we will focus on 11:60750247 SNP (found during Individual View with DaTscan).

First, we search of the SNP is mapped within or close to any gene. To do this let's use 
UCSC genome browser (https://genome.ucsc.edu/).

Our SNP (rs79761481 with RsID) maps within CD6 gene! That is a good starting point. 
We now explore if there are any gene interacting with CD6.
Let's recover CD6 PPI!

To do this we will use STRINGdb R package:
```{r}
string_db <- STRINGdb$new( 
  version = "11", species = 9606
)
protein <-  string_db$mp("CD6")
neig <- string_db$get_neighbors(protein)
string_db$plot_network(neig)
```

As we can see CD6 gene has many direct and indirect interactors. Let's the retrieve 
their names:

```{r}
neig.genes <- sapply(
  neig, 
  function(x){
    base <- "https://string-db.org/api/tsv/get_string_ids?"
    id <- paste0("identifiers=", x, "%0dcdk2&")
    species <-  "species=9606"
    url <- paste0(base,id,species)
    res <- GET(url)
    con <- rawToChar(res$content)
    df  <- read.table(text = con, sep = "\t", header = TRUE)
    return(df$preferredName[1])
  }
)
names(neig.genes) <- NULL
neig.genes
```

Now that we have CD6 interactors names, we can search the pathways in which they participate.
Let's go to Reactiome database (https://reactome.org/).

Once we downloaded the list of CD6 interactors pathways, we can intersect the recovered pathways
with those found during our enrichment analyses.

```{r}
files = list.files(path="../tutorialICHI/resultsRNAseq/DATSCAN", pattern="*.csv", full.names = T)
files = c(files, list.files(path="../tutorialICHI/resultsRNAseq/MRI", pattern="*.csv", full.names = T) )

for (react in c("/resultsRNAseq/CD6_PPI_Reactome_Pathways.csv")){
  cat("******* PPI of: ", strsplit(react,"_")[[1]][1]," *******")
  reactome = read.csv(paste(getwd(),react, sep = ""), header=T, sep=",")
  reactome = reactome[reactome$Entities.FDR<0.1,]
  if(nrow(reactome)==0){break}
  for (i in files){
   pheno = strsplit(i,"_")[[1]][2]
   cat("\nPheno: ", pheno, "\n")
   df = read.csv(i,sep="\t")[,c(1,8)]
   print((intersect(reactome$Pathway.name,df$pathway)))
   cat("\n\n")
 }
}

```

We found 13 pathways in common!

Therefore, we can conclude that SNP rs79761481 can potentially perturb these pathways and
this pobservation has also been confirmed by transcriptomic data.

Moreover, the found pathways have been linked to PD in literature (Kannarkat GT *et al.*, Journal of Parkinson's disease 2013; Xie L *et al.*, Journal of neuroinflammation 2010; Hu Y *et al.*, Frontiers in neurology 2020).