3_Enrichment.Rmd

---
title: "Enrichment Analysis"
author: "Steve Pederson"
date: "`r format(Sys.time(), '%d %B, %Y')`"
output: workflowr::wflow_html
editor_options:
  chunk_output_type: console
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(
    autodep = TRUE,
    echo = TRUE,
    warning = FALSE,
    message = FALSE,
    fig.align = "center"
)
```


# Setup

```{r loadPackages}
library(tidyverse)
library(magrittr)
library(edgeR)
library(scales)
library(pander)
library(goseq)
library(msigdbr)
library(AnnotationDbi)
library(RColorBrewer)
```

```{r setOpts}
theme_set(theme_bw())
panderOptions("table.split.table", Inf)
panderOptions("table.style", "rmarkdown")
panderOptions("big.mark", ",")
```



```{r samplesAndLabels}
samples <- read_csv("data/samples.csv") %>%
  distinct(sampleName, .keep_all = TRUE) %>%
  dplyr::select(sample = sampleName, sampleID, genotype) %>%
  mutate(
    genotype = factor(genotype, levels = c("WT", "Het", "Hom")),
    mutant = genotype %in% c("Het", "Hom"),
    homozygous = genotype == "Hom"
  )
genoCols <- samples$genotype %>%
  levels() %>%
  length() %>%
  brewer.pal("Set1") %>%
  setNames(levels(samples$genotype))
```


```{r loadFits}
dgeList <- read_rds("data/dgeList.rds")
fit <- read_rds("data/fit.rds")
entrezGenes <- dgeList$genes %>%
  dplyr::filter(!is.na(entrezid)) %>%
  unnest(entrezid) %>%
  dplyr::rename(entrez_gene = entrezid)
```

```{r formatP}
formatP <- function(p, m = 0.0001){
  out <- rep("", length(p))
  out[p < m] <- sprintf("%.2e", p[p<m])
  out[p >= m] <- sprintf("%.4f", p[p>=m])
  out
}
```


# Introduction

Enrichment analysis for this datast present some challenges.
Despite normalisation to account for gene length and GC bias, some appeared to still be present in the final results.
In addition, the confounding of incomplete rRNA removal with genotype may lead to other distortions in both DE genes and ranking statistics.

Two methods for enrichment analysis will be undertaken.

1. Testing for enrichment within discrete sets of DE genes as defined in the previous steps
2. Testing for enrichment within ranked lists, regardless of DE status or statistical significance.

Testing for enrichment *within discrete gene sets* will be performed using `goseq` as this allows for the incorporation of a single covariate as a predictor of differential expression.
GC content, gene length and correlation with rRNA removal can all be supplied as separate covariates.

Testing for enrichment *with ranked lists* will be performed using `fry` as this can take a dgeList directly.
However, in order for testing to be robust in the context of bias, normalised counts after CQN will be used for generation of a new dgeList for testing.

# Databases used for testing

Data was sourced using the `msigdbr` package.
The initial database used for testing was the Hallmark Gene Sets, with mappings from gene-set to EntrezGene IDs performed by the package authors. 

# Discrete DE Gene Sets

```{r topTables}
topTables <- list.files(
    path = "output", pattern = "Vs.+csv", full.names = TRUE
) %>%
    sapply(read_csv, simplify = FALSE) %>%
    set_names(basename(names(.))) %>%
    set_names(str_remove(names(.), ".csv")) %>%
    lapply(function(x){
        x %>%
            mutate(
                entrezid = dgeList$genes$entrezid[gene_id]
            )
    })
```


## Effects of the *psen2^S4Ter^ mutation

The first step of analysis using `goseq`, regardless of the geneset, is estimation of the probability weight function (PWF) which quantifies the probablity of a gene being considered as DE based on a single covariate.
Two possible covariates (GC content & Gene Length) were checked


```{r pwf}
pwf <- c("gc_content", "length") %>% 
    sapply(
        function(x){
            topTables$psen2VsWT %>%
                dplyr::select(gene_id, DE, covariate = one_of(x)) %>%
                distinct(gene_id, .keep_all = TRUE) %>% 
                with(
                    nullp(
                        DEgenes = structure(
                            as.integer(DE), names = gene_id
                        ), 
                        genome = "danRer10", 
                        id = "ensGene", 
                        bias.data = covariate,
                        plot.fit = FALSE
                    )
                )
        },
        simplify = FALSE
    )
```

```{r checkBias, fig.width=8, fig.height=4, fig.cap = "*Bias data based on each of the possible covariates. Both appeared to impact the probability of a gene being DE to some extent, with gene length being selected as the covariate for usage inthe goseq model.*"}
par(mfrow = c(1, length(pwf)))
names(pwf) %>%
    lapply(function(x){
        plotPWF(pwf[[x]], main = x)
    })
par(mfrow = c(1, 1))
```


### Hallmark Gene Sets


```{r hm}
hm <- msigdbr("Danio rerio", category = "H")  %>% 
  left_join(entrezGenes) %>%
  dplyr::filter(!is.na(gene_id)) 
```


```{r hmByGene}
hmByGene <- hm %>%
  split(f = .$gene_id) %>%
  lapply(extract2, "gs_name")
```

Mappings are required from gene to pathway, and Ensembl identifiers were used to map from gene to pathway, based on the mappings in the previously used annotations (Ensembl Release 96).
A total of `r length(hmByGene)` Ensembl IDs were mapped to pathways from the hallmark gene sets.
This contrasts with `r length(unique(hm$entrez_gene))` EntrezGene IDs.

```{r hmGoSeq}
hmGoseq <- goseq(pwf$length, gene2cat = hmByGene) %>%
    as_tibble %>%
  dplyr::filter(numDEInCat > 0) %>%
  mutate( 
    adjP = p.adjust(over_represented_pvalue, method = "bonf"),
    FDR = p.adjust(over_represented_pvalue, method = "fdr")
  ) %>%
  dplyr::select(-contains("under"))
```

No hallmark genesets were considered to be enriched after adjusting for the observed length bias in the data.

### KEGG Gene Sets

The same mapping process was applied to KEGG gene sets.

```{r kg}
kg <- msigdbr("Danio rerio", category = "C2", subcategory = "CP:KEGG")  %>% 
  left_join(entrezGenes) %>%
  dplyr::filter(!is.na(gene_id)) 
```


```{r kgByGene}
kgByGene <- kg  %>%
  split(f = .$gene_id) %>%
  lapply(extract2, "gs_name")
```


A total of `r length(kgByGene)` Ensembl IDs were mapped to pathways from the KEGG gene sets.
This contrasts with `r length(unique(kg$entrez_gene))` EntrezGene IDs.

```{r kgGoseq}
kgGoseq <- goseq(pwf$length, gene2cat = kgByGene) %>%
  as_tibble %>%
  dplyr::filter(numDEInCat > 0) %>%
  mutate( 
    adjP = p.adjust(over_represented_pvalue, method = "bonf"),
    FDR = p.adjust(over_represented_pvalue, method = "fdr")
  ) %>%
  dplyr::select(-contains("under"))
```

```{r kgGoseqTable}
kgGoseq %>%
  dplyr::slice(1:5) %>%
  dplyr::rename(p = over_represented_pvalue) %>%
  mutate(
    p = formatP(p),
    adjP = formatP(adjP),
    FDR = formatP(FDR)
  ) %>%
  dplyr::select(
    Category = category,
    DE = numDEInCat,
    `Set Size` = numInCat,
    p,
    `p~bonf~` = adjP,
    `p~FDR~` = FDR
  ) %>%
  pander(
    justify = "lrrrrr",
    caption = paste(
      "The", nrow(.), "most highly-ranked KEGG pathways.",
      "Bonferroni-adjusted p-values are the most stringent and give high",
      "confidence when below 0.05.",
    )
  )
```

Notably, the KEGG gene-set for Ribosomal genes was detected as enriched. 
This is likely to be due to the previosuly discussed contaminant.

### GO Gene Sets

The same mapping process was applied to GO gene sets.

```{r goSummaries}
goSummaries <- url("https://uofabioinformaticshub.github.io/summaries2GO/data/goSummaries.RDS") %>%
  readRDS() %>%
  mutate(
    Term = Term(id),
    gs_name = Term %>% str_to_upper() %>% str_replace_all("[ -]", "_"),
    gs_name = paste0("GO_", gs_name)
    )
```


```{r go}
go <- msigdbr("Danio rerio", category = "C5") %>% 
  left_join(entrezGenes) %>%
  dplyr::filter(!is.na(gene_id)) 
```


```{r goByGene}
minPath <- 3
goByGene <- go %>%
  left_join(goSummaries) %>% 
  dplyr::filter(shortest_path >= minPath) %>%
  split(f = .$gene_id) %>%
  lapply(extract2, "gs_name")
```


A total of `r length(goByGene)` Ensembl IDs were mapped to pathways from the GO gene sets.
This contrasts with `r length(unique(go$entrez_gene))` EntrezGene IDs. 
In addition GO genesets were restricted to those with `r minPath` or more steps back to each ontology root.

```{r goGoseq}
goGoseq <- goseq(pwf$length, gene2cat = goByGene) %>%
  as_tibble %>%
  dplyr::filter(numDEInCat > 0) %>%
  mutate( 
    adjP = p.adjust(over_represented_pvalue, method = "bonf"),
    FDR = p.adjust(over_represented_pvalue, method = "fdr")
  ) %>%
  dplyr::select(-contains("under"))
```

```{r goGoseqTable}
goGoseq %>%
  dplyr::filter(adjP < 0.05) %>%
  dplyr::rename(p = over_represented_pvalue) %>%
  mutate(
    p = formatP(p),
    adjP = formatP(adjP),
    FDR = formatP(FDR)
  ) %>%
  dplyr::select(
    Category = category,
    DE = numDEInCat,
    `Set Size` = numInCat,
    p,
    `p~bonf~` = adjP,
    `p~FDR~` = FDR
  ) %>%
  pander(
    justify = "lrrrrr",
    caption = paste(
      "The", nrow(.), "most highly-ranked GO terms.",
      "Bonferroni-adjusted p-values are the most stringent and give high",
      "confidence when below 0.05, and all terms reached this threshold.",
      "However, most terms indicate the presence of rRNA once again."
    )
  )
```


## Effects of Homozygous Vs Heterozygous mutants

As only a small number of genes were detected as being classified as DE for this comparison, no testing was performed on this as a discrete set of DE genes.