separation_data_analysis.Rmd

---
title: "Separation RNA-seq analysis"
author: "Liza Brusman"
date: "2024-10-30"
output: github_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE, warning = FALSE)
```

```{r, message=FALSE}
library(tidyverse)
library(ggplot2)
library(ggfortify)
library(ggrepel)
library(GGally)
library(wesanderson)
library(viridis)
library(readxl)
library(car)
library(gridExtra)
library(ez)
library(ggpubr)
library(DESeq2)
library(EnhancedVolcano)
library(RRHO2)
library(ggpubr)
library(UpSetR)
library(scales)
library(stringi)
library(SuperExactTest)
library(ComplexHeatmap)
library(colorRamp2)
library(ggridges)
library(vioplot)
source("separation_data_fxns.R")
```

import sequencing data and associated metadata
```{r}
#import metadata
metrics <- read_excel("../../docs/Input_RNAseq_metrics.xlsx", sheet = "Correct order prelim")
metrics <- mutate_at(metrics, vars("Animal", "Cohort", "Pairing", "Timepoint", "Separation", "ProcessingGroup", "PPrefCat", "DissectionOrder", "CageChange", "ParentCode", "PairingAgeCat", "SACAgeCat", "RIN", "InputRNACat", "TRAP", "InputRNA"), as.factor)
#remove animal 2945 that didn't have a partner preference
metrics <- metrics %>% filter(Animal != 2945)

#import RNA-seq data
inputs <- read.csv("../../docs/merged_all_inputs.txt", stringsAsFactors=FALSE)
inputs <- inputs %>% filter_all(all_vars(.>10)) %>% distinct(genes, .keep_all = TRUE) %>% filter(!is.na(genes)) %>% column_to_rownames("genes")
#remove animal 2945
inputs <- inputs %>% select(-c(Counts2945I))
```

first, filter for only cohorts i care about (cohorts analogous to snRNA-seq experiment)
```{r}
want_anis <- metrics$Animal[metrics$Cohort %in% c("SS48P", "SS4I", "OS48P", "OS4I")] %>% unfactor()
want_idxs <- match(want_anis, metrics$Animal)

want_inputs <- inputs[,want_idxs]
want_metrics <- metrics %>% filter(Animal %in% want_anis)
```

import module genes
```{r}
mod_genes <- read.csv("G:/My Drive/pateiv/snRNAseq_voles/src/hotspot/docs/new_clusts_hotspot-gene-modules.csv")
all_genes <- intersect(mod_genes$Gene, rownames(want_inputs))
#get just module-6 genes
mod6_genes <- mod_genes$Gene[mod_genes$Module == 6]
mod6_genes <- intersect(mod6_genes, rownames(want_inputs))
mod_inputs <- want_inputs[all_genes,]
```

get normalized counts
```{r}
norm_cts_all <- get_norm_cts(want_inputs, want_metrics)
```

data wrangling to get dfs that i can use for heatmaps, etc.
```{r}
#transpose normalized counts
mod_inputs_meta <- transpose_norm_cts(norm_cts_all)

#normalize counts (z-score)
norm_new <- zscore_norm_cts(mod_inputs_meta, all_genes)

#get group mean of z-scored counts
norm_new_summ_mtx <- avg_zscored_cts(norm_new)
```

first, let's check to see if any animals are outliers based on raw counts and z-scored counts \
look at histogram of z-scored counts per animal (and raw counts per animal if you want) \
animal 2599 appears to be an outlier based on the distribution of z-scored counts (a very wide distribution compared to the others)
```{r}
#for zscored counts
for (ani in norm_new$Animal) {
  print(ani)
  ani_df <- norm_new %>% filter(Animal == ani)
  ani_cts <- ani_df[,-c(1:3)]
  ani_cts <- t(ani_cts)
  hist(ani_cts[,1], xlim = c(-4, 4), ylim = c(0, 800), main = paste0(ani, " ", ani_df$Cohort))
}

# #for raw counts
# for (ani in colnames(mod_inputs)) {
#   ani_df <- mod_inputs %>% select(ani)
#   ani_cts <- ani_df[1]
#   hist(ani_cts[,1], xlim = c(0, 200000), main = ani)
# }
```

animals with medium partner preference scores: "Counts2599I", "Counts2887P", "Counts2758P", "Counts2647I" \
excluding animal 2599
```{r}
#filter out animal 2599
want_inputs_excl <- want_inputs %>% select(-c("Counts2599I")) #"Counts2599I" , "Counts2887I", "Counts2758I", "Counts2647I"
want_metrics_excl <- want_metrics %>% filter(!Animal %in% c("2599")) #, "2887", "2758", "2647"

norm_cts_all_excl <- get_norm_cts(want_inputs_excl, want_metrics_excl)

mod_inputs_meta_excl <- transpose_norm_cts(norm_cts_all_excl)

#normalize counts (z-score)
norm_new_excl <- zscore_norm_cts(mod_inputs_meta_excl, all_genes)

norm_new_summ_mtx_excl <- avg_zscored_cts(norm_new_excl)
```

```{r, fig.width = 7, fig.height = 22}
#to make and save heatmaps for each module
mods <- 6 #numbers here indicate the modules you want to make heatmap  (all modules would be 1:23)
for (m in mods) {
  make_heatmap(norm_new_summ_mtx_excl, mod = m, save = FALSE)
}

```

to get out clustering from heatmap and make ridgeline plots of z-scored expression for each Module-6 gene
```{r, fig.width = 7, fig.height = 22}
mods <- 6 #numbers here indicate the modules you want to make heatmap of
for (m in mods) {
  make_ridgelines(norm_new_summ_mtx_excl, norm_new, norm_cts_all_excl, metrics, mod = m, save = FALSE)
}
```

this block is to look at the difference between TRUE distribution of gene change scores (per module) vs. the shuffled distribution (animal IDs shuffled). \
first, i calculated the median for each gene in "paired" and "separated" animals, then subtracted these values to find a change score \
then, i shuffled animal ID to create a null distribution of change scores (should be centered around zero) \
then, i used an F-test to compare these distributions \
if the module is truly "socially sensitive", the variance of the TRUE change scores should be greater than the variance of the shuffled change scores \
here i have combined OS and SS to JUST look at paired vs. isolated
```{r, warning = FALSE, eval = FALSE}
##for non-shuffled change scores - only looking at separation condition
combine_norm_new <- mod_inputs_meta_excl %>% select(c("Animal", "Separation", all_genes)) #mod_inputs_meta_excl
combine_norm_new[all_genes] <- scale(combine_norm_new[all_genes])

#take median value of each group (i.e. separation condition)
combine_OSSS <- combine_norm_new %>% select(c("Animal", "Separation", all_genes)) %>% group_by(Separation) %>% dplyr::summarise_if(is.numeric, median)
rownames(combine_OSSS) <- combine_OSSS$Separation
stash_names <- rownames(combine_OSSS)

#do some data wrangling to get the isolated - paired value for each gene
combine_OSSS_mat <- combine_OSSS[,-1]
combine_OSSS_mat <- t(combine_OSSS_mat) %>% as.data.frame()
colnames(combine_OSSS_mat) <- stash_names
combine_OSSS_mat$IsoMinPair <- combine_OSSS_mat$Isolated - combine_OSSS_mat$Paired
combine_OSSS_mat$Gene <- rownames(combine_OSSS_mat)
combine_OSSS_mat <- combine_OSSS_mat %>% merge(mod_genes, on = "Gene")

#set up a df to hold p-values from F-test to compare distributions
hold_comb_values <- data.frame()
#we need to iterate through seeds for the shuffle because there aren't that many animals, so the seed can make a big difference
for (s in 1:100) {
  print(s)
  set.seed(s)
  
  #shuffle animal ID and then do same data manipulation as above
  norm_cts_shuff <- norm_cts_all_excl
  colnames(norm_cts_shuff) <- sample(colnames(norm_cts_shuff))
  
  mod_inputs_meta_shuff <- transpose_norm_cts(norm_cts_shuff)

  combine_norm_new_shuff <- mod_inputs_meta_shuff %>% select(c("Animal", "Separation", all_genes))
  combine_norm_new_shuff[all_genes] <- scale(combine_norm_new_shuff[all_genes])
  
  combine_OSSS_shuff <- combine_norm_new_shuff %>% select(c("Animal", "Separation", all_genes)) %>% group_by(Separation) %>% dplyr::summarise_if(is.numeric, median)
  rownames(combine_OSSS_shuff) <- combine_OSSS_shuff$Separation
  stash_names <- rownames(combine_OSSS_shuff)
  
  combine_OSSS_mat_shuff <- combine_OSSS_shuff[,-1]
  combine_OSSS_mat_shuff <- t(combine_OSSS_mat_shuff) %>% as.data.frame()
  colnames(combine_OSSS_mat_shuff) <- stash_names
  combine_OSSS_mat_shuff$IsoMinPair <- combine_OSSS_mat_shuff$Isolated - combine_OSSS_mat$Paired
  combine_OSSS_mat_shuff$Gene <- rownames(combine_OSSS_mat_shuff)
  combine_OSSS_mat_shuff <- combine_OSSS_mat_shuff %>% merge(mod_genes, on = "Gene")
  
  #iterate through modules to do a separate stats test for each seed (true vs. shuffled). this will give you 100 p-values per module
  for (mod in 1:23) { 
    mod_df <- combine_OSSS_mat %>% filter(Module == mod)
    mod_df_shuff <- combine_OSSS_mat_shuff %>% filter(Module == mod)
    
    all_var_test <- var.test(mod_df$IsoMinPair, mod_df_shuff$IsoMinPair)
    
    ##you can plot as a split violin if you want, but probably don't want to make 2300 plots
    # vioplot(IsoMinPair~Module, data=mod_df, col = "palevioletred", plotCentre = "line", side = "left", ylim = c(-3, 3))
    # vioplot(IsoMinPair~Module, data=mod_df_shuff, col = "lightblue", plotCentre = "line", side = "right", ylim = c(-3, 3), add = T, areaEqual = F)

    mini_df <- data.frame(Module = mod, seed = s, pval = all_var_test[3]$p.value, Fstat = all_var_test$statistic)
  
    hold_comb_values <- rbind(hold_comb_values, mini_df)
  }
}
```

SAVE csv
```{r, eval = FALSE}
setwd("output/")
write.csv(hold_comb_values, "shuffled_seed_module_combineOSSS_pvals_100iter_excl2599.csv")
```

read back in csv
```{r}
hold_comb_values <- read.csv("output/shuffled_seed_module_combineOSSS_pvals_100iter_excl2599.csv")
```

get summary stats from these tests
```{r}
iter_summ_comb <- hold_comb_values %>% group_by(Module) %>% summarise(mean_pval = mean(pval), mean_F = mean(Fstat), median_pval = median(pval))
```

plot the data! \
first plot is bubble plot where size is % times p-value is under cutoff, mean/median pval under cutoff is color \
second plot is a boxplot of the log10(p-values) for each module
```{r}
hold_comb_values$log10pval <- log10(hold_comb_values$pval)

#get percent of iterations for each module with log10(p-val) < -5, then mean and median p-val
iter_summ_comb_cutoff <- hold_comb_values %>% group_by(Module) %>% summarise(perc_iters = sum(log10pval < -5), mean_pval = mean(log10pval), median_pval = median(log10pval))

#set up color palette for plotting mean p-values
myPalette <- colorRampPalette(c("whitesmoke", "#67001F"))(200) #, "#f0f921"


#plot bubble plot
p <- ggplot(iter_summ_comb_cutoff, aes(x = as.factor(Module), y = 1, fill = abs(mean_pval))) +
  geom_point(aes(size = perc_iters), alpha = 1, shape = 21) +
  scale_fill_gradientn(colors = myPalette, limits = c(0, max(abs(iter_summ_comb_cutoff$mean_pval))),
                       guide = guide_colorbar(frame.colour = "black", ticks.colour = "black")) +
  # scale_size_continuous(limits = c(0, min(iter_summ_comb_cutoff$mean_pval))) +
  # scale_fill_manual(values = clust_cols, guide = FALSE) +
  # scale_color_manual(values = clust_cols, guide = FALSE) +
  scale_y_discrete(labels = 1:23) +
  theme(panel.background = element_blank(), panel.border = element_rect(colour = "grey70", fill = NA, size = 1))

print(p)

#plot boxplot
p2 <- ggplot(hold_comb_values, aes(x = as.factor(Module), y = log10(pval))) + geom_boxplot() + geom_hline(yintercept = -5, color = "red", linetype = "dashed") + theme_classic()
print(p2)

#pvalue plot
setwd("output/")
ggsave("bubbleplot_combineSSOS_pvals_cutoff_Ftests_excl2599.pdf", p)
ggsave("boxplot_combineSSOS_pvals_Ftests_excl2599.pdf", p2)

```
heatmap clustering animals by separation condition
```{r, fig.height = 8, fig.width = 10}
setwd("output/")
mods <- 6
for (mod in mods) {
  m_genes <- mod_genes$Gene[mod_genes$Module == mod]
  m_genes <- intersect(m_genes, rownames(norm_cts_all_excl))
  
  norm_cts_mod <- norm_new_excl %>% select(Animal, Cohort, Separation, m_genes)
  rownames(norm_cts_mod) <- norm_cts_mod$Animal
  col_fun = colorRamp2(c(min(norm_cts_mod[,-c(1:3)]), 0, max(norm_cts_mod[,-c(1:3)])), c("#053061","white", "#67001F"))
  
  # row_ha <- rowAnnotation(Cohort = norm_cts_mod$Cohort, col = list(Cohort = c("OS48P" = "deeppink3", "SS48P" = "darkgreen", "OS4I" = "lightpink", "SS4I" = "lightgreen")))
  row_ha <- rowAnnotation(Separation = norm_cts_mod$Separation, col = list(Separation = c("Paired" = "darkblue", "Isolated" = "lightblue")))
  ht <- Heatmap(norm_cts_mod[,-c(1:3)], right_annotation = row_ha, column_title = paste0("Module-", mod), col = col_fun) #km = 4, 
  draw(ht)
  #get out dendrogram info
  og_dend <- row_dend(ht)
  #to save heatmap
  pdf(paste0("module_", mod, "_heatmap_paired_isolated_excl2599.pdf"))
  draw(Heatmap(norm_cts_mod[,-c(1:3)], right_annotation = row_ha, cluster_rows = og_dend, split = 4, column_title = paste0("Module-", mod), col = col_fun)) #km = 4,
  dev.off()

}
```


```{r}
sessionInfo()
```