AnalyzingMultipleDatasets.Rmd

---
title: "Analyzing Multiple Datasets"
package: BRGenomics
output:
  BiocStyle::html_document:
    toc: true
    toc_float: true
  BiocStyle::pdf_document:
    toc: true
vignette: |
  %\VignetteIndexEntry{Analyzing Multiple Datasets}
  %\VignetteEncoding{UTF-8}
  %\VignetteEngine{knitr::rmarkdown}
---

# Lists

To efficiently work with several datasets, we recommend storing the GRanges 
objects within a standard, named list, e.g. `grl <- list(a_rep1 = gr1, b_rep1 = 
gr2, ...)`.

```{r, message = FALSE}
library(BRGenomics)
data("PROseq")
data("txs_dm6_chr4")
```

```{r}
# make 3 datasets
ps1 <- PROseq[seq(1, length(PROseq), 3)]
ps2 <- PROseq[seq(2, length(PROseq), 3)]
ps3 <- PROseq[seq(3, length(PROseq), 3)]

# use the "=" assignment in list() to give names to the list elements
ps_list <- list(ps1 = ps1, ps2 = ps2, ps3 = ps3)
names(ps_list)
ps_list
```

A named list like the one above can be passed as an argument to nearly every 
function in BRGenomics, and many functions will automatically return 
dataframes, or melted dataframes that use the list names as the sample names 
(which can simplify plotting with `ggplot2` or `lattice`).

---

_Note that BRGenomics does also support the use of `GRangesList` or
`CompressedGRangesList` classes for grouping multiple datasets. However, care
should be taken when using these, as many functions with methods for `GRanges`
objects also have methods for `GRangesList` objects. Furthermore, `GRangesList` 
objects can be automatically coerced into `CompressedGRangesList` objects, 
which can lower memory usage but can also incur significant performance 
penalties._

---

## Example: Summarizing Signal from Multiple Samples

We can pass our list of GRanges directly to `getCountsByRegions` to 
simultaneously count reads for each dataset, and melt the result for plotting 
with `ggplot`:

```{r}
getCountsByRegions(ps_list, txs_dm6_chr4[1:5], ncores = 1)
```

```{r}
# melt, and use the optional region_names argument
txs_counts <- getCountsByRegions(ps_list, txs_dm6_chr4, melt = TRUE,
                                 region_names = txs_dm6_chr4$tx_name,
                                 ncores = 1)
head(txs_counts)
```

```{r}
library(ggplot2)
ggplot(txs_counts, aes(x = sample, y = signal, fill = sample)) + 
  geom_violin() + 
  theme_bw()
```

## Example: Making Browser Shots in R

By using the `getCountsByPositions()` on several datasets over a single 
region-of-interest, we can bake our own genome browser shots within R. Using 
the same region we plotted earlier:

```{r}
cbp_maxtx <- getCountsByPositions(ps_list, txs_dm6_chr4[135], 
                                  melt = TRUE, ncores = 1)
head(cbp_maxtx)
```

```{r, fig.height=4, fig.width=6}
ggplot(cbp_maxtx, aes(x = position, y = signal)) + 
    facet_wrap(~sample, ncol = 1, strip.position = "right") + 
    geom_col(size = 0.5, color = "darkgray") +
    coord_cartesian(expand = FALSE) +
    labs(title = txs_dm6_chr4$tx_name[135],
         x = "Distance from TSS", y = "PRO-seq Signal") + 
    theme_classic() + 
    theme(strip.text.y = element_text(angle = 0),
          strip.background = element_blank(),
          axis.line.x = element_blank(),
          axis.ticks.x = element_blank())
```

# Multiplexed GRanges

There is another data structure that is widely supported with BRGenomics, 
called a __multiplexed GRanges__. A multiplexed GRanges is a single GRanges 
object that contains multiple metadata fields containing basepair-resolution 
coverage data for different datasets. We currently recommend users use lists of 
GRanges objects, but you might find that multiplexed GRanges objects have 
performance benefits for your data (notes on this below).

A multiplexed GRanges object can be made using the `mergeGRangesData()` with 
option `multiplex = TRUE`.

```{r}
ps_multi <- mergeGRangesData(ps1, ps2, ps3, multiplex = TRUE, ncores = 1)
ps_multi
```

As in any GRanges object, the metadata fields are accessible as a dataframe 
with the `mcols()` function, and individual columns are accessible with the `$` 
operator.^[The dataframe, by default, is not a base R `data.frame`, but rather 
an S4 `DataFrame`. The distinction isn't important for end users, and it's 
unlikely users will encounter any reason to coerce the class using 
`as.data.frame`.]

```{r}
mcols(ps_multi)
```

```{r, collapse = TRUE}
ps_multi$ps1[1:5]
```

This data structure can be passed to most functions in BRGenomics. To be 
explicit, users should set the `field` argument, when it exists, to set the 
datasets for which calculations should be performed:

```{r}
# for all datasets (all fields), get counts in the first 5 transcripts
getCountsByRegions(ps_multi, txs_dm6_chr4[1:5], 
                   field = names(mcols(ps_multi)),
                   ncores = 1)

# get counts for ps2 dataset only
getCountsByRegions(ps_multi, txs_dm6_chr4[1:5], 
                   field = "ps2", ncores = 1)

# if no field is given, most functions will default to using all fields
getCountsByRegions(ps_multi, txs_dm6_chr4[1:5], 
                   ncores = 1)
```

In our experience with PRO-seq data, and to our surprise, a list of GRanges 
objects typically outperforms multiplexed GRanges on a typical laptop, despite 
the theoretical benefits of multiplexing. 

---

_In principle, the multiplexed GRanges structure is designed to reduce memory 
usage and increase performance, but this has not been our experience when 
dealing with sparse, basepair-resolution data like PRO-seq. Overlapping signal 
with regions of interest (e.g. in `getCountsByRegions()` or 
`getCountsByPosition()`) is relatively efficient for reasonably sized GRanges 
objects, and these calculations scale reasonably well with multicore 
processing. While multiplexing means that signal overlapping only has to be 
performed once, we've found that this is a relatively small benefit in practice 
that is easily offset by having to do signal calculations on large sparse 
vectors of signal counts. However, the relative benefits of using lists or 
multiplexing are likely to depend on the nature of the data being analyzed, as 
well as well as computer hardware._

---

# Saving Binary R Files

While the handling of GRanges data in BRGenomics is relatively fast, the 
initial importation of bigWig or bedGraph files as GRanges objects remains a 
noticeable bottleneck. This may be tolerable for interactive workflows, in 
which data is imported once before undergoing lengthy analysis, but the 
bottleneck is a significant detriment for users who regularly import data.

To avoid this bottleneck, users can save reusable data structures as binary R 
files, which effectively save the memory state of R objects. Not only are these 
objects rapidly reloaded into memory upon importation, but they have the added 
benefit of saving the user from repeating data formatting.

Any R object can be saved to storage using the `saveRDS()` function, and 
re-imported using `readRDS()`. 

```{r, eval = FALSE}
# save the PRO-seq GRanges for later import
saveRDS(PROseq, file = "~/PROseq.RData") 

# save a list of GRanges
saveRDS(ps_list, file = "~/ps_list.RData")

# re-import
ps_list <- readRDS("~/ps_list.RData")
```

The `save()` and `load()` commands can also be used to accomplish the same 
thing, although they work slightly differently.^[Unlike the `saveRDS()`/
`readRDS()` commands, the `read()`/`load()` commands maintain the original name
of the objects. For instance, if you `save(PROseq, file = "~/ps.RData")`, and 
in a new R session run `read("~/ps.RData")`, a new object called `PROseq` will 
be created in your new environment. (Note that this is the same way that 
RStudio saves your current working environment to disk, i.e. it saves the 
entire environment into an RData file, which can then be reloaded, remaking 
every data object).]