2_Co-expression_network_characterisation.Rmd

## Co-expression network characterisation (Fig. 1e)


### Core- and tissue-specific co-expression

We characterised our tissue-specific co-expression networks based on GTEx. Our hypothesis is that genes that are highly co-expression across all tissues are likely required for cellular developments and survival, and should show a strong correlation with of essentiality. In this analysis, we downloaded the list of human essential genes from the [OGEE database (v2)]( http://ogee.medgenius.info), also included in `/data/OGEE_esential_genes_20190416.txt`. 


```{r}
## coexpresssion - share edges
## goal: to observe whether the shared edges among co-expression networks are essential

# 0 - load required data
## load coexel sum
library(pacman)
p_load(tidyverse, cowplot, knitr)

coex_el_sum_grouped = readRDS("../cache/coexpression_edge_counts_by_group.RDS")

# create a vector of total probability for each class
coex_el_sum_score  = coex_el_sum_grouped %>% 
  ungroup() %>% 
  group_by(essential_edge_score) %>% 
  summarise(count = sum(n)) %>% pull(count, name = essential_edge_score)

coex_el_sum_grouped %>% 
  dplyr::rename(n_tissues = n_binned_relabel) %>%
  group_by(n_tissues) %>%
  summarise(n_edges = sum(n), percent =  sum(n)*100/sum(coex_el_sum_score)) %>%
  kable
```

The plot for higher essentiality aas number of genes increased is shown below.

```{r bar plot essentiality per tissue pinning}
bar_essential = ggplot(coex_el_sum_grouped) + 
  geom_bar(aes(x = n_binned_relabel, y = n, fill = score), stat = "identity",  position="fill") + 
  xlab("Number of tissues") + ylab("Edge proportion") + theme_cowplot() + 
  theme(legend.position = "bottom", axis.text.x = element_text(size = 10)) + 
  guides(fill = guide_legend(title = "Essential gene in edge")) + scale_fill_manual(values = c('grey60','#9ecae1','#3182bd'))

bar_essential
#ggsave("./Figs/essentiality_co-expression.pdf", bar_essential, width = 4, height = 4)
```

```{r}
coex_el_sum = readRDS("../cache/coexpression_raw_edge_counts.RDS") %>% ungroup
coex_el_sum_by_tissue <- coex_el_sum %>% 
  count(n, essential_edge_score, name = "count")
```




```{r edge count in absolute value, include=FALSE, eval=FALSE}
## save absolute values for number of edges in each bin
coex_bin_count = coex_el_sum_grouped %>% 
  group_by(n_binned_relabel) %>% 
  dplyr::rename(count = n)

EE_count <- coex_el_sum %>% ungroup() %>% count(n_binned, essential_edge_score)


coex_bin_count_plot = ggplot(coex_bin_count) + geom_bar(aes(x = n_binned_relabel, y = count), stat = "identity", alpha = 0.7) +
  xlab("Tissues") + ylab("edge count") + theme_cowplot() #+ scale_y_log10() 

coex_bin_count_plot
```


```{r assessment, include=FALSE, eval=FALSE}
#### Perform hypergeometric distribution of the combination
coex_bin_both_essential = coex_el_sum_grouped %>% filter(essential_edge_score==2) %>% select(n_binned_relabel, n)
coex_bin_not_essential = coex_el_sum_grouped %>% filter(essential_edge_score<2) %>% group_by(n_binned_relabel) %>% summarise(n_non = sum(n))
coex_bin_pval = inner_join(coex_bin_both_essential, coex_bin_not_essential, by = "n_binned_relabel")
coex_bin_pval$n_one = coex_el_sum_grouped %>% filter(essential_edge_score>0) %>% group_by(n_binned_relabel) %>% summarise(n_one = sum(n)) %>% pull(n_one)
#all_genes  = unique(c(levels(coex_el_sum$A), levels(coex_el_sum$B)))
#n_essential = sum(all_genes %in% essential_genes_sym)
# use binomial distribution to estimate the significance
allessentialpairs = coex_el_sum %>% filter(essential_edge_score == 2) %>% nrow()
oneessentialpairs = coex_el_sum %>% filter(essential_edge_score == 1) %>% nrow()
allpairs = nrow(coex_el_sum)
p_essential = allessentialpairs/allpairs
p_oneessential = (oneessentialpairs+allessentialpairs)/allpairs

coex_bin_pval = coex_bin_pval %>% rowwise %>% mutate(pval = pbinom(n, n+n_non, p_essential, lower.tail = F, log.p = T),
                                                     ratio = n/(n+n_non),
                                                     oddratio = ratio/p_essential,
                                                     ratio_one = n_one/(n+n_non),
                                                     oddratio_one = ratio_one/p_oneessential)

ggplot(coex_bin_pval %>% ungroup) + geom_point(aes(x=n_binned_relabel, y = oddratio))


## roc
coex_el_sum = coex_el_sum %>% arrange(-n, -essential_edge_score)
coex_el_sum$rank = 1:nrow(coex_el_sum)
coex_el_sum= coex_el_sum %>% mutate(bothEssential = essential_edge_score ==2,
                                    atLeastOneEssential = essential_edge_score > 0)
#coex_el_sum %>% bothEssentialCumSum = cumsum(coex_el_sum$bothEssential)

library(pROC)
coex_el_roc  = coex_el_sum_grouped %>%  group_by(n_binned_relabel) %>% tally()

coex_el_roc  = coex_el_sum %>% group_by(n_binned) %>% arrange(-n) %>% mutate(EdgeRank = (1:n()),
                                                                             EdgeRankProp = EdgeRank/n(),
                                                                             bothEssentialCumSum = cumsum(bothEssential),
                                                                             #  TrueProb = (bothEssentialCumSum)/sum(),
                                                                             FPR = (EdgeRank - bothEssentialCumSum)/(max(EdgeRank)-max(bothEssentialCumSum)),
                                                                             TPR = (bothEssentialCumSum/max(bothEssentialCumSum))) %>% 
  select(n_binned, EdgeRank, EdgeRankProp, FPR, TPR) #%>% filter(!duplicated(n))


coex_el_roc  = coex_el_sum %>% ungroup %>% arrange(-n) %>%
  mutate(EdgeRank = (1:nrow(coex_el_roc)),
         EdgeRankProp = EdgeRank/nrow(coex_el_roc),
         bothEssentialCumSum = cumsum(bothEssential),
         #  TrueProb = (bothEssentialCumSum)/sum(),
         FPR = (EdgeRank - bothEssentialCumSum)/(max(EdgeRank)-max(bothEssentialCumSum)),
         TPR = (bothEssentialCumSum/max(bothEssentialCumSum))) %>% 
  select(EdgeRank, EdgeRankProp, FPR, TPR) #%>% filter(!duplicated(n))


coex_el_roc_sampled =coex_el_roc %>% group_by(n_binned) %>% dplyr::sample_n(1000)
coex_el_roc_sampled =coex_el_roc %>% dplyr::sample_frac(0.0005)
#roc1 = ggplot(coex_el_roc, aes(factor(n, levels = rev(1:53), ordered = T), maxval)) + geom_point() + theme_cowplot()
roc1 = ggplot(coex_el_roc_sampled, aes(FPR, TPR)) + geom_line(aes(col = factor(n_binned, ordered = T))) + theme_cowplot() +
  xlab("number of tissues found") + ylab("Pboth edges are essential")

roc1 = ggplot(coex_el_roc_sampled, aes(FPR, TPR)) + geom_line() + theme_cowplot() +
  xlab("number of tissues found") + ylab("Pboth edges are essential")

ggsave("./analysis/network_characterisation/essentiality/ROC_essential.pdf", roc1)

###################
## essentiality and rare disease genes
rarediseasegenes = process_disease_genes_data("./Disease_gene_assoc/Orphanet/output/table_disease_gene_assoc_orphanet.tsv", 20, 2000)
alldiseasegenes = unique(unlist(rarediseasegenes$disgene_list))
table(alldiseasegenes %in% essential_genes_sym)

disease_gene_essential = sapply(rarediseasegenes$disgene_list, function(x) table(x %in% essential_genes_sym))
disease_gene_essential = melt(disease_gene_essential)
colnames(disease_gene_essential) = c("essential", "count", "diseasename")

disease_gene_essential_prop = disease_gene_essential %>% group_by(diseasename) %>% spread(essential, count) %>% mutate(sum = `FALSE`+`TRUE`,
                                                                                                                       propTrue = `TRUE`/sum) %>% filter(sum > 20)

bar_essential = ggplot(disease_gene_essential %>% filter(count > 20)) + geom_bar(aes(diseasename, count, fill = essential), stat = "identity") + scale_y_log10() + coord_flip()+ theme_cowplot()
ggsave("./analysis/network_characterisation/essentiality/essential_by_disease.pdf", bar_essential)

#######################
# add the hypergeometric test to the results

coex_el_sum_stat = coex_el_sum_grouped %>% group_by(n_binned) %>% summarise(n_edge = sum(n))

# map the number of essential-essential edges
coex_el_sum_stat = left_join(coex_el_sum_stat, 
                             coex_el_sum_grouped %>% dplyr::filter(score == 2) %>% ungroup %>% dplyr::select(n_binned, n))

n_essential = sum(unique(unique(coex_el_sum$A), unique(coex_el_sum$B)) %in% essential_genes_sym)
n_nonessential = sum(!unique(unique(coex_el_sum$A), unique(coex_el_sum$B)) %in% essential_genes_sym)

# possibiility of two essential genes connecting to each other (EE), essential-nonessential (NE), nonessential-nonessential (NN)
#n_EE = n_essential * (n_essential-1) / 2
#n_NN = n_nonessential * (n_nonessential-1) / 2
#n_NE = n_essential * n_nonessential

# computed based on what in the data
n_EE = sum(coex_el_sum$essential_edge_score == 2)
n_NN = sum(coex_el_sum$essential_edge_score == 0)
n_NE = sum(coex_el_sum$essential_edge_score == 1)

# compute the hypergeometric pval
binom_p = n_EE/(n_NE+n_NN)
coex_el_sum_stat = coex_el_sum_stat %>% rowwise() %>%
  mutate(hyp_pval = phyper(n, n_EE, (n_NN + n_NE) - (n_edge - n), n_edge, lower.tail = F, log.p = F),
         bin_pval = binom.test(n, n_edge, binom_p, alternative = "greater")$p.value)


```


```{r assessment2, include=FALSE, eval=FALSE}
# Feb 2021
pval_test_from_obs = function(group, method = "binomial"){
  # this counts from the population
  if(method == "binomial"){
    counts = coex_el_sum_grouped %>% dplyr::filter(n_binned == group) %>% pull(n, name = score)
    allcounts = sum(counts)
    p_EE_expected = coex_el_sum_score['2']/sum(coex_el_sum_score)
    testresult <- binom.test(counts['2'], allcounts, p_EE_expected, alternative = "greater")
  }
  return(testresult)
}

result_node_binom_from_obs = sapply(unique(coex_el_sum_grouped$n_binned), function(x) pval_test_from_obs(x))


#=======================
#2 - incorporate essential genes data
source("../functions/fn_source.R")
essential_genes = read_tsv("../data/OGEE_esential_genes_20190416.txt")
essential_genes$geneSymbol = IDconvert(essential_genes$locus, from = "ENSEMBL", to = "SYMBOL")
essential_genes_sym = essential_genes %>% filter(essential == "E") %>% pull(geneSymbol) %>% unique()


# new implementation
pval_test = function(group, method = "binomial"){
  # group
  if(method == "binomial"){
    sub_edges = coex_el_sum[coex_el_sum$n_binned==group,]
    sub_nodes = sub_edges %>% select(A,B) %>% as.list() %>% unlist() %>% unique()
    
    N = length(sub_nodes)
    nE = sum(sub_nodes %in% essential_genes_sym) # number of essential genes
    nN = N - nE
    
    nEE = sum(sub_edges$essential_edge_score == 2)  # number of positive drawn = sum of EE edges
    nALL = nrow(sub_edges)
    p_EE_expected = (nE*(nE-1))/(N*(N-1))
    testresult <- binom.test(nEE, nALL, p_EE_expected, alternative = "greater")
  }
  return(c(essential_nodes = nE, all_nodes = N, testresult))
}


result = sapply(unique(coex_el_sum_grouped$n_binned), function(x) pval_test(x))



#=========

# new implementation
pval_test_node = function(group, method = "binomial"){
  print(group)
  all_genes = unique(c(levels(coex_el_sum$A), levels(coex_el_sum$B)))
  all_essential_genes = sum(all_genes %in% essential_genes_sym)
  p_essential = all_essential_genes/length(all_genes)
  # group
  if(method == "binomial"){
    sub_edges = coex_el_sum[coex_el_sum$n_binned==group,]
    sub_nodes = sub_edges %>% select(A,B) %>% as.list() %>% unlist() %>% unique()
    
    N = length(sub_nodes)
    nE = sum(sub_nodes %in% essential_genes_sym) # number of essential genes

    testresult <- binom.test(nE, N, p_essential, alternative = "greater")
  }
  if(method == "hypergeom"){
    sub_edges = coex_el_sum[coex_el_sum$n_binned==group,]
    sub_nodes = sub_edges %>% select(A,B) %>% as.list() %>% unlist() %>% unique()
    
    N = length(sub_nodes)
    nE = sum(sub_nodes %in% essential_genes_sym) # number of essential genes
    
    testresult <- phyper(q = nE, m = all_essential_genes, n = length(all_genes) - all_essential_genes, k = N, lower.tail = F)
  }
  return(c(essential_nodes = nE, all_nodes = N, testresult))
}


result_node_binom = sapply(unique(coex_el_sum_grouped$n_binned), function(x) pval_test_node(x))
result_node_phyper = sapply(unique(coex_el_sum_grouped$n_binned), function(x) pval_test_node(x, method = "hypergeom"))





```