variance_components.Rmd

---
title: "Variance components analysis"
author: "Yuanhua Huang & Davis J. McCarthy"
site: workflowr::wflow_site
---

## Load libraries 

```{r, warning=FALSE, message=FALSE}
library(viridis)
library(tidyverse)
library(ggpubr)
library(ggrepel)
library(cowplot)
library(reshape2)
```

## Results aggregated across all lines

Previously, we conducted variance components analysis combining scRNA-seq data
across all lines. We used the `variancePartion` package, fitting assigned 
clone, line, plate (on which cells were processed) and cellular detection rate
(cdr; proportion of genes per cell with non-zero expression) as random effects.

The output of the model gives us the proportion of variance explained, for 
each gene, by clone, line, plate, cdr and residuals. 

```{r}
varPart.df <- as.data.frame(
  read.table("data/variance_components/fit_per_gene_highVar.csv", sep = ","))
colnames(varPart.df) <- c("clone", "line", "plate", "cdr", "residuals")
varPart.df$gene_id <- rownames(varPart.df)
varPart.df <- as_data_frame(varPart.df)
head(varPart.df)
```

We ran the analysis for `r nrow(varPart.df)` highly variable genes (as 
identified using methods in the `scran` package).

We detect `r sum(varPart.df$clone > 0.05)` genes with more than 5% of variance
explained by clone and `r sum((varPart.df$clone > varPart.df$plate) & (varPart.df$clone > 0.01))` genes for which variance explained by clone is both
greater than 1% and also greater than the variance explained by plate (a good
proxy for technical effects on variability in gene expression between cells).


```{r}
varPart.mel <- tidyr::gather(varPart.df, key = "variable", value = "value", 
                             -gene_id)
```

```{r}
dat_dir <- "data/variance_components/donorVar/"

fig.violin <- varPart.mel %>%
  dplyr::mutate(variable = replace(variable, variable == "cdr", 
                                   "cell. det. rate")) %>%
  dplyr::mutate(variable = factor(variable, 
                                  levels = c("residuals", "cell. det. rate",
                                             "plate", "line", "clone"))) %>%
  ggplot(aes(x = variable, y = 100 * value)) + 
  geom_violin(aes(fill = variable), scale = "width") + 
  geom_boxplot(width = 0.07, fill = "grey", outlier.colour = "black") + 
  ylab("Variance explained (%)") + 
  xlab("") + 
  scale_fill_manual(values = c("gray70", "#f7efe2", "#f9a603", "#f25c00",
                               "#f70025")) +
  coord_flip() +
  theme(legend.position = "none") +
  theme(plot.title = element_text(hjust = 0.5))

ggsave("figures/variance_components/varpart_violin_alllines.png", 
       fig.violin, height = 4, width = 6.5, dpi = 300)
ggsave("figures/variance_components/varpart_violin_alllines.pdf", 
       fig.violin, height = 4, width = 6.5, dpi = 300)
ggsave("figures/variance_components/varpart_violin_alllines_skinny.png", 
       fig.violin, height = 4, width = 4.5, dpi = 300)
ggsave("figures/variance_components/varpart_violin_alllines_skinny.pdf", 
       fig.violin, height = 4, width = 4.5, dpi = 300)

fig.violin
```


```{r}
# plotVarPart( varPart_mat )
# idx_order <- order(varPart_mat$clone, decreasing = TRUE)
# plotPercentBars( varPart_mat[idx_order[1:10],] )
```


### Plot line component against clone component

```{r, warning=FALSE, message=FALSE}
varPart.df %>%
  ggplot(aes(x = 100 * line, y = 100 * clone)) +
  geom_point(alpha = 0.5) +
  geom_smooth(colour = "firebrick", alpha = 0.7) +
  coord_cartesian(ylim = c(0, 25)) +
  xlab("Variance explained by line (%)") +
  ylab("Variance explained by clone (%)")
```


## Results for individual lines

### Load individual lines

We load a SingleCellExperiment object for each line containing expression data,
cell metadata and gene metadata.

```{r, message=FALSE, warning=FALSE}
params <- list()
params$callset <- "filt_lenient.cell_coverage_sites" 
lines <- c("euts", "fawm", "feec", "fikt", "garx", "gesg", "heja", "hipn", 
            "ieki", "joxm", "kuco", "laey", "lexy", "naju", "nusw", "oaaz", 
            "oilg", "pipw", "puie", "qayj", "qolg", "qonc", "rozh", "sehl", 
            "ualf", "vass", "vuna", "wahn", "wetu", "xugn", "zoxy")
## Load SCE objects
sce_unst_list <- list()
for (don in lines) {
    sce_unst_list[[don]] <- readRDS(file.path("data/sces/", 
        paste0("sce_", don, "_with_clone_assignments.", params$callset, ".rds")))
    cat(paste("reading", don, ":   ", ncol(sce_unst_list[[don]]), "cells\n"))
}
```

We load variance component analysis results for each line.

```{r}
varPart_list <- list()
for (i in seq_along(lines)) {
  df <- read.csv(paste0("data/variance_components/donorVar/",
                        lines[i], ".var_part.var1.csv"))
    colnames(df) <- c("clone", "plate", "cell. det. rate", "residuals")
    df$gene_id <- rownames(df)
    df$line <- lines[i]
    varPart_list[[i]] <- as_data_frame(df)
}
names(varPart_list) <- lines
```

For each line we thus have a data.frame with the variance components results.

```{r}
head(varPart_list[[1]])
```

### Number of genes with substantial clone component

We observe larger clone effects if we fit the linear mixed model for each line
separately. A large number of genes for each line have >5% of variance explained
by clone (median: `r median(sapply(varPart_list, function(x) sum(x$clone > 0.05)))`).

```{r}
sort(sapply(varPart_list, function(x) sum(x$clone > 0.05)))
```


### Plot results for individual lines

For each line we will plot the variance explained by clone for the to 400 genes
with most variance explained by clone (for that line).

First, we need to process the raw variance component results from each line.

```{r}
n_top <- 400
n_lines <- length(varPart_list)

n_hv_genes <- rep(NA, n_lines)
n_cells_use <- rep(NA, n_lines)
n_genes_use <- rep(NA, n_lines)
var_top_mean <- rep(NA, n_lines)
var_top_list <- list()
for (i in seq_len(length(varPart_list))) {
    sort_idx <- order(varPart_list[[i]]$clone, decreasing = TRUE)
    var_top_list[[i]] <- varPart_list[[i]][sort_idx[1:n_top], "clone", 
                                           drop = FALSE]
    var_top_mean[i] <- mean(var_top_list[[i]]$clone, na.rm = TRUE)
    n_hv_genes[i] <- sum(varPart_list[[i]]$clone > 0.25, na.rm = TRUE)
    min_var <- 1.0
    cell_idx <- which(sce_unst_list[[lines[i]]]$assigned != "unassigned")
    n_cells_use[i] <- length(cell_idx)
    n_genes_use[i] <- nrow(varPart_list[[i]])
}
don_sort_idx <- order(var_top_mean, decreasing = TRUE)

df.line <- data_frame(line_id = lines, 
                       clone_cells = n_cells_use, 
                       n_genes_use = n_genes_use, 
                       n_hv_genes = n_hv_genes, 
                       var_top_mean = var_top_mean)[don_sort_idx, ]
df.line <- cbind(rank = seq_len(nrow(df.line)), df.line)

var_full <- c()
idx_full <- c()
don_full <- c()
rank_full <- c()
for (ii in don_sort_idx) {        
    var_full <- c(var_full, var_top_list[[ii]]$clone)
    idx_full <- c(idx_full, seq_len(n_top))
    don_full <- c(don_full, rep(lines[ii], n_top))
    rank_full <- c(rank_full, rep(ii, n_top))
}

line_info <- rep("other", length(don_full))
median_line <- lines[don_sort_idx[round(length(don_sort_idx)/2)]]
# line_info[don_full == median_line] <- "median"
line_info[don_full == "joxm"] <- "joxm"
# print(median_line)
# summary(n_hv_genes)

df <- data_frame(varPart = var_full, gene_rank = idx_full, 
                 lines = don_full, line_rank = rank_full,
                 line_info = line_info)
rm(line_info)

df <- df %>%
  group_by(gene_rank) %>%
  summarise(varPart = median(varPart)) %>%
  dplyr::mutate(lines = "median", line_info = "median", line_rank = NA) %>%
  bind_rows(., df)

```

The table below provides a summary of the per-line variance component results.
We can read off the mean proportion of variance explained by clone for the top
400 genes for each line (`var_top_mean`), as well as other information for each
line.

```{r}
df.line
```

Now we can plot a curve for each line showing the variance explained by clone
for the top 400 genes.

```{r, warning=FALSE, message=FALSE}
df_labs <- dplyr::filter(df, gene_rank == 200) %>%
  dplyr::mutate(labs = ifelse(varPart > 0.33, lines, ""))

fig.curve <- ggplot(df, aes(colour = line_info)) +
  geom_line(aes(x = gene_rank, y = 100 * varPart, color = line_info, 
                size = line_info, group = lines)) +
  scale_color_manual(values = c("firebrick", "black", "gray60"),
                     name = "line") +
  scale_size_manual(values = c(0.5, 1, 0.5), name = "line") +
  geom_line(data = df[df$lines == "joxm", ],
            aes(x = gene_rank, y = 100 * varPart), size = 0.5, color = "firebrick") +
  geom_line(data = df[df$lines == "median", ],
            aes(x = gene_rank, y = 100 * varPart), size = 1, color = "black") +
  geom_text(aes(x = gene_rank - 15, y = 100 * varPart, label = labs),
                   nudge_y = 0, nudge_x = 0, 
                   data = df_labs, colour = "gray40", size = 2) +
  xlab("Top genes") + ylab("Variance explained by clone (%)") +
  ggtitle("") + xlim(0, 400) +
  theme(legend.position = c(0.8, 0.8))

ggsave("figures/variance_components/varpart_curve_perline.png", 
       fig.curve, height = 4, width = 6.5, dpi = 300)
ggsave("figures/variance_components/varpart_cruve_perline.pdf", 
       fig.curve, height = 4, width = 6.5, dpi = 300)
ggsave("figures/variance_components/varpart_curve_perline_skinny.png", 
       fig.curve +  theme(legend.position = c(0.7, 0.8)),
       height = 4, width = 4.5, dpi = 300)
ggsave("figures/variance_components/varpart_cruve_perline_skinny.pdf", 
       fig.curve +  theme(legend.position = c(0.7, 0.8)), 
       height = 4, width = 4.5, dpi = 300)

fig.curve
```

Each line has a substantial number of genes with a substantial proportion of 
variance explained by clone. For example, the number of genes with more than
25% of variance explained by line across lines:

```{r}
sapply(varPart_list, function(x) sum(x$clone > 0.25))
```

There are a large number of genes in each line with more than 5% of variance
explained by clone:

```{r}
sapply(varPart_list, function(x) sum(x$clone > 0.05))
```



## Plots including line selection information

We have inferred selection dynamics from somatic variants detected from 
whole-exome sequencing data. We can overlay selection status on the plot above 
to look for any relationship between selection and variance explained by clone
in top genes by line.

```{r}
line_selected <- c("euts", "fawm", "fikt", "garx", "pipw", "puie", "qolg", "rozh")
line_neutral <- c("hipn", "nusw", "qonc", "sehl", "ualf", "xugn")
line_undetermined <- c("feec", "gesg", "heja", "ieki", "joxm", "kuco", "laey", 
                        "lexy", "naju", "oaaz", 
                        "oilg", "qayj", "vass", "vuna", "wahn", "wetu", "zoxy")
df$selection <- "undetermined"
df$selection[df$lines %in% line_neutral] <- "neutral"
df$selection[df$lines %in% line_selected] <- "selection"
df$selection[df$lines %in% "median"] <- "median"
df$selection <- factor(df$selection, 
                       levels = c("neutral", "undetermined", "selection", 
                                  "median"))
```


```{r}
df_labs <- dplyr::filter(df, gene_rank == 200) %>%
  dplyr::mutate(labs = ifelse(varPart > 0.33, lines, ""))

fig.curve <- ggplot(df, aes(colour = selection)) +
  geom_line(aes(x = gene_rank, y = 100 * varPart, color = selection, 
                size = selection, group = lines)) +
  scale_color_manual(values =  c("dodgerblue", "#CCCCCC", "dodgerblue4", 
                                 "black"), name = "selection\ndynamics") +
  scale_size_manual(values = c(0.5, 0.5, 0.5, 1), name = "selection\ndynamics") +
  geom_line(data = df[df$lines == "median", ],
            aes(x = gene_rank, y = 100 * varPart), size = 1, color = "black") +
  geom_text(aes(x = gene_rank - 15, y = 100 * varPart, label = labs),
                   nudge_y = 0, nudge_x = 0, 
                   data = df_labs, colour = "gray40", size = 2) +
  xlab("Top genes") + ylab("Variance explained by clone (%)") +
  ggtitle("") + xlim(0, 400) +
  theme(legend.position = c(0.8, 0.8))

fig.curve
```

There is not any obvious relationship between selection dynamics and variance
explained by clone here.

## Write mean variance explained to file

Write a table with the mean fraction of variance explained by clone for the top
400 genes per line out to file.

```{r}
idx <- order(var_top_mean, decreasing = TRUE)
line_var_df <- data.frame(line = lines[idx], 
                           meanFraction400 = var_top_mean[idx])
write.csv(line_var_df, "output/variance_components/line_var_top400.csv")
```