01-preprocessing.Rmd

---
title: "Pre-processing"
---

```{r knitr, include = FALSE}
DOCNAME = "01-preprocessing"
NOW <- Sys.time()

# Time chunks during knitting
knitr::knit_hooks$set(timeit = function(before) {
    if (before) {
        print(paste("Start:", Sys.time()))
        NOW <- Sys.time()
    } else {
        print(paste("Stop:", Sys.time()))
        print(Sys.time() - NOW)
    }
})  

knitr::opts_chunk$set(
    autodep        = TRUE,
    cache          = TRUE,
    cache.path     = paste0("cache/", DOCNAME, "/"),
    cache.comments = FALSE,
    echo           = TRUE,
    error          = FALSE,
    fig.align      = "center",
    fig.width      = 10,
    fig.height     = 8,
    message        = FALSE,
    warning        = FALSE,
    timeit         = TRUE
)
```

```{r libaries, cache = FALSE}
# scRNA-seq
library("DropletUtils")
library("SingleCellExperiment")

# Tidyverse
library("tidyverse")
```

```{r source, cache = FALSE}
source(here::here("R/load.R"))
source(here::here("R/annotate.R"))
source(here::here("R/output.R"))
```

```{r bpparam, cache = FALSE}
bpparam <- BiocParallel::MulticoreParam(workers = 10)
```

Introduction
============

In this document we are going to read in the unfiltered counts matrix produced
by `Cell Ranger` and determine which of those droplets contain cells using the
`DropletUtils` package.

```{r load}
path <- here::here("data/cellranger")
raw <- read10x(path, dataset = "Orgs123")

filt_barcodes <- readr::read_lines(file.path(path, "filtered_barcodes.tsv.gz"))

colData(raw)$CellRangerFilt <- paste(colData(raw)$Barcode, 
                                     colData(raw)$Sample, sep = "-") %in%
    filt_barcodes
```

The raw dataset has `r nrow(raw)` features and `r ncol(raw)` droplets.

Barcode ranks
=============

```{r barcodes}
empty_thresh <- 100

bc_ranks <- barcodeRanks(counts(raw), lower = empty_thresh)

colData(raw)$BarcodeRank   <- bc_ranks$rank
colData(raw)$BarcodeTotal  <- bc_ranks$total
colData(raw)$BarcodeFitted <- bc_ranks$fitted

bc_data <- colData(raw) %>%
    as.data.frame() %>%
    select(Cell, Kept = CellRangerFilt, Rank = BarcodeRank,
           Total = BarcodeTotal, Fitted = BarcodeFitted) %>%
    arrange(Rank)
```

Let's start by ordering the droplets according to their total counts and
plotting this on a log scale. This let's us see the distribution of total
counts.

```{r barcodes-plot}
ggplot(bc_data, aes(x = Rank, y = Total)) +
    geom_point(shape = 1, aes(colour = Kept)) +
    geom_line(aes(y = Fitted), colour = "red") +
    geom_hline(yintercept = bc_ranks$knee,
               colour = "dodgerblue", linetype = "dashed") +
    annotate("text", x = 0, y = bc_ranks$knee, label = "Knee",
             colour = "dodgerblue", hjust = 0, vjust = -1) +
    geom_hline(yintercept = bc_ranks$inflection,
               colour = "forestgreen", linetype = "dashed") +
    annotate("text", x = 0, y = bc_ranks$inflection, label = "Inflection",
             colour = "forestgreen", hjust = 0, vjust = -1) +
    geom_hline(yintercept = empty_thresh,
               colour = "darkorchid", linetype = "dashed") +
    annotate("text", x = 0, y = empty_thresh, label = "Empty threshold",
             colour = "darkorchid", hjust = 0, vjust = -1) +
    scale_x_log10(labels = scales::number) +
    scale_y_log10(labels = scales::number) +
    scale_colour_manual(values = c("black", "violet")) +
    ylab("Total counts") +
    theme_minimal()
```

This is typical of what we see for 10x experiment where there is a sharp drop
off between droplets with lots of counts and those without many. The inflection
and knee points are methods for identifying the transition between
distributions. These are roughly associated with the cells selected by `Cell
Ranger` The empty threshold line indicates the point at which we assume droplets
must be empty (total counts <= `r empty_thresh`).

Empty drops
===========

We will now look at identifying which droplets to select using the `emptyDrops`
method. This method tests whether the composition of a droplet is significantly
different from the ambient RNA in the sample which is obtained by pooling the
empty droplets. Droplets with very large counts are also automatically retained.

```{r empty-drops}
set.seed(1)

emp_iters <- 30000

emp_drops <- emptyDrops(counts(raw), lower = empty_thresh, niters = emp_iters,
                        test.ambient = TRUE, BPPARAM = bpparam)
```

`emptyDrops` calculates p-values using a permutation approach. Let's check that
we are usually a sufficient number of iterations. If there are any droplets
that have non-significant p-values but are limited by the number of permuations
the number should be increased.

```{r empty-drops-iters}
emp_fdr <- 0.01
is_cell <- emp_drops$FDR <= emp_fdr
is_cell[is.na(is_cell)] <- FALSE

colData(raw)$EmpDropsLogProb <- emp_drops$LogProb
colData(raw)$EmpDropsPValue  <- emp_drops$PValue
colData(raw)$EmpDropsLimited <- emp_drops$Limited
colData(raw)$EmpDropsFDR     <- emp_drops$FDR
colData(raw)$EmpDropsFilt    <- is_cell

table(Limited = emp_drops$Limited, Significant = is_cell)
```

Another way to check the `emptyDrops` results is to look at the droplets below
our empty threshold. We are assuming that these droplets only contain ambient
RNA and therefore the null hypothesis should be true and the distribution of
p-values should be approximately uniform.

```{r empty-drops-pvals}
colData(raw) %>%
    as.data.frame() %>%
    filter(BarcodeTotal <= empty_thresh,
           BarcodeTotal > 0) %>%
    ggplot(aes(x = EmpDropsPValue)) +
    geom_histogram() +
    xlab("p-value") +
    theme_minimal()
```

Peaks near zero would tell us that not all of the droplets below the threshold
are truly empty and that we should lower it. 

We can also plot the negative log-probability against the total counts to see
which droplets `emptyDrops` has selected.

```{r empty-drops-plot}
colData(raw) %>%
    as.data.frame() %>%
    filter(!is.na(EmpDropsFilt)) %>%
    ggplot(aes(x = BarcodeTotal, y = -EmpDropsLogProb, colour = EmpDropsFilt)) +
    geom_point() +
    scale_colour_discrete(name = "Significant") +
    xlab("Total counts") +
    ylab("-log(probability)") +
    theme_minimal()
```

There were `r sum(emp_drops$Total <= empty_thresh)` droplets with less than
`r empty_thresh` counts which were used to make up the ambient RNA pool. Of the
remaining `r sum(emp_drops$Total > empty_thresh)` droplets `r sum(is_cell)` were
found to have profiles significantly different from the ambient RNA and should
contain cells.

Comparison
==========

Let's quickly compare differences between the default we get from `Cell Ranger`
and what we have got from using `DropletUtils`. First let's look at the number
of cells identified by each method.

```{r compare-total}
colData(raw) %>%
    as.data.frame() %>%
    group_by(Sample) %>%
    summarise(`Cell Ranger` = sum(CellRangerFilt),
              emptyDrops    = sum(EmpDropsFilt, na.rm = TRUE)) %>%
    gather(key = Method, value = Total, -Sample) %>%
    ggplot(aes(x = Method, y = Total, fill = Sample)) +
    geom_col() +
    ylab("Number of cells") +
    theme_minimal()
```

From this plot we can see that `emptyDrops` has identified many more cells and
that they seem to be evenly spread across the three samples. Well will also
check to see that the methods are finding the same cells.

```{r compare-overlap}
plot_data <- colData(raw) %>%
    as.data.frame() %>%
    mutate(`Cell Ranger only` =  CellRangerFilt & !EmpDropsFilt,
           `emptyDrops only`  = !CellRangerFilt &  EmpDropsFilt,
           Both               =  CellRangerFilt &  EmpDropsFilt) %>%
    group_by(Sample) %>%
    summarise(`Cell Ranger only` = sum(`Cell Ranger only`),
              `emptyDrops only`  = sum(`emptyDrops only`, na.rm = TRUE),
              Both               = sum(Both, na.rm = TRUE)) %>%
    gather(key = Selected, value = Count, -Sample) %>%
    mutate(Selected = factor(Selected,
                             levels = c("Both", "Cell Ranger only",
                                        "emptyDrops only")))

ggplot(plot_data, aes(x = Selected, y = Count, fill = Sample)) +
    geom_col() +
    theme_minimal()

plot_data %>%
    group_by(Selected) %>%
    summarise(Total = sum(Count))
```

Most cells are identified by both methods with `emptyDrops` identifying many
more additional cells. There are only a few cells that are identified only by
Cell Ranger. We will perform further quality control of cells so at this stage
we will keep cells that are selected by either method.

```{r select}
selected <- raw[, colData(raw)$CellRangerFilt | colData(raw)$EmpDropsFilt]

colData(selected)$SelMethod <- "Both"
colData(selected)$SelMethod[!colData(selected)$CellRangerFilt] <- "emptyDrops"
colData(selected)$SelMethod[!colData(selected)$EmpDropsFilt] <- "CellRanger"
```

Annotation
==========

Now that we have a dataset that just contains actual cells we will add some
extra annotation. This includes downloading feature annotation from `BioMart`
and assigning cell cycle stages using the `cyclone` function in the `scran`
package as well as calculating a range of quality control metrics using
`scater`.

```{r annotate}
selected <- annotateSCE(selected,
                        org        = "human",
                        add_anno   = TRUE,
                        host       = "jul2018.archive.ensembl.org",
                        calc_qc    = TRUE,
                        calc_cpm   = TRUE,
                        cell_cycle = TRUE,
                        BPPARAM    = bpparam)

glimpse(as.data.frame(colData(selected)))
glimpse(as.data.frame(rowData(selected)))
```

Summary
=======

Parameters
----------

This table describes parameters used and set in this document.

```{r parameters}
params <- list(
     list(
        Parameter = "empty_thresh",
        Value = empty_thresh,
        Description = "Droplets with less than this many counts are empty"
    ),
    list(
        Parameter = "emp_iters",
        Value = emp_iters,
        Description = "Number of iterations for emptyDrops p-values"
    ),
    list(
        Parameter = "emp_fdr",
        Value = emp_fdr,
        Description = "FDR cutoff for emptyDrops"
    ),
    list(
        Parameter = "n_cells",
        Value = ncol(selected),
        Description = "Number of cells selected"
    )
)

metadata(selected)$Params <- list()
metadata(selected)$Params[DOCNAME] <- params

params <- jsonlite::toJSON(params, pretty = TRUE)
knitr::kable(jsonlite::fromJSON(params))
```

Output files
------------

This table describes the output files produced by this document. Right click
and _Save Link As..._ to download the results.

```{r save}
write_rds(selected, here::here("data/processed/01-selected.Rds"))
```

```{r output}
dir.create(here::here("output", DOCNAME), showWarnings = FALSE)

readr::write_lines(params, here::here("output", DOCNAME, "parameters.json"))

knitr::kable(data.frame(
    File = c(
        getDownloadLink("parameters.json", DOCNAME)
    ),
    Description = c(
        "Parameters set and used in this analysis"
    )
))
```

Session information
-------------------

```{r session-info, cache = FALSE}
devtools::session_info()
```