Data analysis involves a large amount of janitor work – munging and cleaning data to facilitate downstream data analysis. This chapter demonstrates techniques for advanced data manipulation and analysis with the split-apply-combine strategy. We will use the dplyr package in R to effectively manipulate and conditionally compute summary statistics over subsets of a “big” dataset containing many observations.
This chapter assumes a basic familiarity with R (Chapter 1) and data frames (Chapter 2).
Recommended reading: Review the Introduction (10.1) and Tibbles vs. data.frame (10.3) sections of the R for Data Science book. We will initially be using the read_* functions from the readr package. These functions load data into a tibble instead of R’s traditional data.frame. Tibbles are data frames, but they tweak some older behaviors to make life a little easier. These sections explain the few key small differences between traditional data.frames and tibbles.
3.1 Review
3.1.1 Our data
We’re going to use the yeast gene expression dataset described on the data frames chapter in Chapter 2. This is a cleaned up version of a gene expression dataset from Brauer et al. Coordination of Growth Rate, Cell Cycle, Stress Response, and Metabolic Activity in Yeast (2008) Mol Biol Cell 19:352-367. This data is from a gene expression microarray, and in this paper the authors are examining the relationship between growth rate and gene expression in yeast cultures limited by one of six different nutrients (glucose, leucine, ammonium, sulfate, phosphate, uracil). If you give yeast a rich media loaded with nutrients except restrict the supply of a single nutrient, you can control the growth rate to any rate you choose. By starving yeast of specific nutrients you can find genes that:
Raise or lower their expression in response to growth rate. Growth-rate dependent expression patterns can tell us a lot about cell cycle control, and how the cell responds to stress. The authors found that expression of >25% of all yeast genes is linearly correlated with growth rate, independent of the limiting nutrient. They also found that the subset of negatively growth-correlated genes is enriched for peroxisomal functions, and positively correlated genes mainly encode ribosomal functions.
Respond differently when different nutrients are being limited. If you see particular genes that respond very differently when a nutrient is sharply restricted, these genes might be involved in the transport or metabolism of that specific nutrient.
You can download the cleaned up version of the data here. The file is called brauer2007_tidy.csv. Later on we’ll actually start with the original raw data (minimally processed) and manipulate it so that we can make it more amenable for analysis.
3.1.2 Reading in data
We need to load both the dplyr and readr packages for efficiently reading in and displaying this data. We’re also going to use many other functions from the dplyr package. Make sure you have these packages installed as described on the setup chapter (Appendix A).
# Load packageslibrary(readr)library(dplyr)# Read in dataydat<-read_csv(file="data/brauer2007_tidy.csv")# Display the dataydat# Optionally, bring up the data in a viewer window# View(ydat)
The dplyr package is a relatively new R package that makes data manipulation fast and easy. It imports functionality from another package called magrittr that allows you to chain commands together into a pipeline that will completely change the way you write R code such that you’re writing code the way you’re thinking about the problem.
When you read in data with the readr package (read_csv()) and you had the dplyr package loaded already, the data frame takes on this “special” class of data frames called a tbl (pronounced “tibble”), which you can see with class(ydat). If you have other “regular” data frames in your workspace, the as_tibble() function will convert it into the special dplyr tbl that displays nicely (e.g.: iris <- as_tibble(iris)). You don’t have to turn all your data frame objects into tibbles, but it does make working with large datasets a bit easier.
You can read more about tibbles in Tibbles chapter in R for Data Science or in the tibbles vignette. They keep most of the features of data frames, and drop the features that used to be convenient but are now frustrating (i.e. converting character vectors to factors). You can read more about the differences between data frames and tibbles in this section of the tibbles vignette, but the major convenience for us concerns printing (aka displaying) a tibble to the screen. When you print (i.e., display) a tibble, it only shows the first 10 rows and all the columns that fit on one screen. It also prints an abbreviated description of the column type. You can control the default appearance with options:
options(tibble.print_max = n, tibble.print_min = m): if there are more than n rows, print only the first m rows. Use options(tibble.print_max = Inf) to always show all rows.
options(tibble.width = Inf) will always print all columns, regardless of the width of the screen.
3.3 dplyr verbs
The dplyr package gives you a handful of useful verbs for managing data. On their own they don’t do anything that base R can’t do. Here are some of the single-table verbs we’ll be working with in this chapter (single-table meaning that they only work on a single table – contrast that to two-table verbs used for joining data together, which we’ll cover in a later chapter).
They all take a data frame or tibble as their input for the first argument, and they all return a data frame or tibble as output.
3.3.1 filter()
If you want to filter rows of the data where some condition is true, use the filter() function.
The first argument is the data frame you want to filter, e.g. filter(mydata, ....
The second argument is a condition you must satisfy, e.g. filter(ydat, symbol == "LEU1"). If you want to satisfy all of multiple conditions, you can use the “and” operator, &. The “or” operator | (the pipe character, usually shift-backslash) will return a subset that meet any of the conditions.
==: Equal to
!=: Not equal to
>, >=: Greater than, greater than or equal to
<, <=: Less than, less than or equal to
Let’s try it out. For this to work you have to have already loaded the dplyr package. Let’s take a look at LEU1, a gene involved in leucine synthesis.
# First, make sure you've loaded the dplyr packagelibrary(dplyr)# Look at a single gene involved in leucine synthesis pathwayfilter(ydat, symbol=="LEU1")
# Optionally, bring that result up in a View window# View(filter(ydat, symbol == "LEU1"))# Look at multiple genesfilter(ydat, symbol=="LEU1"|symbol=="ADH2")
# Look at LEU1 expression at a low growth rate due to nutrient depletion# Notice how LEU1 is highly upregulated when leucine is depleted!filter(ydat, symbol=="LEU1"&rate==.05)
Let’s look at this graphically. Don’t worry about what these commands are doing just yet - we’ll cover that later on when we talk about ggplot2. Here’s I’m taking the filtered dataset containing just expression estimates for LEU1 where I have 36 rows (one for each of 6 nutrients \(\times\) 6 growth rates), and I’m piping that dataset to the plotting function, where I’m plotting rate on the x-axis, expression on the y-axis, mapping the value of nutrient to the color, and using a line plot to display the data.
Look closely at that! LEU1 is highly expressed when starved of leucine because the cell has to synthesize its own! And as the amount of leucine in the environment (the growth rate) increases, the cell can worry less about synthesizing leucine, so LEU1 expression goes back down. Consequently the cell can devote more energy into other functions, and we see other genes’ expression very slightly raising.
Exercise 1
Display the data where the gene ontology biological process (the bp variable) is “leucine biosynthesis” (case-sensitive) and the limiting nutrient was Leucine. (Answer should return a 24-by-7 data frame – 4 genes \(\times\) 6 growth rates).
Gene/rate combinations had high expression (in the top 1% of expressed genes)? Hint: see ?quantile and try quantile(ydat$expression, probs=.99) to see the expression value which is higher than 99% of all the data, then filter() based on that. Try wrapping your answer with a View() function so you can see the whole thing. What does it look like those genes are doing? Answer should return a 1971-by-7 data frame.
3.3.1.1 Aside: Writing Data to File
What we’ve done up to this point is read in data from a file (read_csv(...)), and assigning that to an object in our workspace (ydat <- ...). When we run operations like filter() on our data, consider two things:
The ydat object in our workspace is not being modified directly. That is, we can filter(ydat, ...), and a result is returned to the screen, but ydat remains the same. This effect is similar to what we demonstrated in our first session.
# Assign the value '50' to the weight object.weight<-50# Print out weight to the screen (50)weight# What's the value of weight plus 10?weight+10# Weight is still 50weight# Weight is only modified if we *reassign* weight to the modified valueweight<-weight+10# Weight is now 60weight
More importantly, the data file on disk (data/brauer2007_tidy.csv) is never modified. No matter what we do to ydat, the file is never modified. If we want to save the result of an operation to a file on disk, we can assign the result of an operation to an object, and write_csv that object to disk. See the help for ?write_csv (note, write_csv() with an underscore is part of the readr package – not to be confused with the built-in write.csv() function).
# What's the result of this filter operation?filter(ydat, nutrient=="Leucine"&bp=="leucine biosynthesis")# Assign the result to a new objectleudat<-filter(ydat, nutrient=="Leucine"&bp=="leucine biosynthesis")# Write that out to diskwrite_csv(leudat, "leucinedata.csv")
Note that this is different than saving your entire workspace to an Rdata file, which would contain all the objects we’ve created (weight, ydat, leudat, etc).
3.3.2 select()
The filter() function allows you to return only certain rows matching a condition. The select() function returns only certain columns. The first argument is the data, and subsequent arguments are the columns you want.
# Select just the symbol and systematic_nameselect(ydat, symbol, systematic_name)
Notice above how the original data doesn’t change. We’re selecting out only certain columns of interest and throwing away columns we don’t care about. If we wanted to keep this data, we would need to reassign the result of the select() operation to a new object. Let’s make a new object called nogo that does not contain the GO annotations. Notice again how the original data is unchanged.
# create a new dataset without the go annotations.nogo<-select(ydat, -bp, -mf)nogo
The mutate() function adds new columns to the data. Remember, it doesn’t actually modify the data frame you’re operating on, and the result is transient unless you assign it to a new object or reassign it back to itself (generally, not always a good practice).
The expression level reported here is the \(log_2\) of the sample signal divided by the signal in the reference channel, where the reference RNA for all samples was taken from the glucose-limited chemostat grown at a dilution rate of 0.25 \(h^{-1}\). Let’s mutate this data to add a new variable called “signal” that’s the actual raw signal ratio instead of the log-transformed signal.
Mutate has a nice little feature too in that it’s “lazy.” You can mutate and add one variable, then continue mutating to add more variables based on that variable. Let’s make another column that’s the square root of the signal ratio.
The arrange() function does what it sounds like. It takes a data frame or tbl and arranges (or sorts) by column(s) of interest. The first argument is the data, and subsequent arguments are columns to sort on. Use the desc() function to arrange by descending.
First, re-run the command you used above to filter the data for genes involved in the “leucine biosynthesis” biological process and where the limiting nutrient is Leucine.
Wrap this entire filtered result with a call to arrange() where you’ll arrange the result of #1 by the gene symbol.
Wrap this entire result in a View() statement so you can see the entire result.
3.3.5 summarize()
The summarize() function summarizes multiple values to a single value. On its own the summarize() function doesn’t seem to be all that useful. The dplyr package provides a few convenience functions called n() and n_distinct() that tell you the number of observations or the number of distinct values of a particular variable.
Notice that summarize takes a data frame and returns a data frame. In this case it’s a 1x1 data frame with a single row and a single column. The name of the column, by default is whatever the expression was used to summarize the data. This usually isn’t pretty, and if we wanted to work with this resulting data frame later on, we’d want to name that returned value something easier to deal with.
# Get the mean expression for all genessummarize(ydat, mean(expression))
# A tibble: 1 × 1
`mean(expression)`
<dbl>
1 0.00337
# Use a more friendly name, e.g., meanexp, or whatever you want to call it.summarize(ydat, meanexp=mean(expression))
# A tibble: 1 × 1
meanexp
<dbl>
1 0.00337
# Measure the correlation between rate and expression summarize(ydat, r=cor(rate, expression))
# A tibble: 1 × 1
r
<dbl>
1 -0.0220
# Get the number of observationssummarize(ydat, n())
# A tibble: 1 × 1
`n()`
<int>
1 198430
# The number of distinct gene symbols in the data summarize(ydat, n_distinct(symbol))
# A tibble: 1 × 1
`n_distinct(symbol)`
<int>
1 4211
3.3.6 group_by()
We saw that summarize() isn’t that useful on its own. Neither is group_by() All this does is takes an existing data frame and coverts it into a grouped data frame where operations are performed by group.
The real power comes in where group_by() and summarize() are used together. First, write the group_by() statement. Then wrap the result of that with a call to summarize().
# Get the mean expression for each gene# group_by(ydat, symbol)summarize(group_by(ydat, symbol), meanexp=mean(expression))
# Get the correlation between rate and expression for each nutrient# group_by(ydat, nutrient)summarize(group_by(ydat, nutrient), r=cor(rate, expression))
This is where things get awesome. The dplyr package imports functionality from the magrittr package that lets you pipe the output of one function to the input of another, so you can avoid nesting functions. It looks like this: |>. You don’t have to load the magrittr package to use it since dplyr imports its functionality when you load the dplyr package.
Here’s the simplest way to use it. Remember the tail() function. It expects a data frame as input, and the next argument is the number of lines to print. These two commands are identical:
Now, think about this for a minute. What if we wanted to get the correlation between the growth rate and expression separately for each limiting nutrient only for genes in the leucine biosynthesis pathway, and return a sorted list of those correlation coeffients rounded to two digits? Mentally we would do something like this:
Take the ydat dataset
thenfilter() it for genes in the leucine biosynthesis pathway
Now compare that with the mental process of what you’re actually trying to accomplish. The way you would do this without pipes is completely inside-out and backwards from the way you express in words and in thought what you want to do. The pipe operator |> allows you to pass the output data frame from one function to the input data frame to another function.
Nesting functions versus piping
This is how we would do that in code. It’s as simple as replacing the word “then” in words to the symbol |> in code. (There’s a keyboard shortcut that I’ll use frequently to insert the |> sequence – you can see what it is by clicking the Tools menu in RStudio, then selecting Keyboard Shortcut Help. On Mac, it’s CMD-SHIFT-M.)
What are the four most highly expressed genes when the growth rate is restricted to 0.05 by restricting glucose? Show only the symbol, expression value, and GO terms. Hint: 4 pipes: filter, arrange, head, and select.
# A tibble: 4 × 4
symbol expression bp mf
<chr> <dbl> <chr> <chr>
1 ADH2 6.28 fermentation* alcohol dehydrogenase activity
2 HSP26 5.86 response to stress* unfolded protein binding
3 MLS1 5.64 glyoxylate cycle malate synthase activity
4 HXT5 5.56 hexose transport glucose transporter activity*
Exercise 5
When the growth rate is restricted to 0.05, what is the average expression level across all genes in the “response to stress” biological process, separately for each limiting nutrient? What about genes in the “protein biosynthesis” biological process? Hint: 3 pipes: filter, group_by, summarize.
Which 10 biological process annotations have the most genes associated with them? What about molecular functions? Hint: 4 pipes: group_by, summarize with n_distinct, arrange, head.
# A tibble: 10 × 2
bp n
<chr> <int>
1 biological process unknown 269
2 protein biosynthesis 182
3 protein amino acid phosphorylation* 78
4 protein biosynthesis* 73
5 cell wall organization and biogenesis* 64
6 regulation of transcription from RNA polymerase II promoter* 49
7 nuclear mRNA splicing, via spliceosome 47
8 DNA repair* 44
9 ER to Golgi transport* 42
10 aerobic respiration* 42
# A tibble: 10 × 2
mf n
<chr> <int>
1 molecular function unknown 886
2 structural constituent of ribosome 185
3 protein binding 107
4 RNA binding 63
5 protein binding* 53
6 DNA binding* 44
7 structural molecule activity 43
8 GTPase activity 40
9 structural constituent of cytoskeleton 39
10 transcription factor activity 38
Exercise 8
How many distinct genes are there where we know what process the gene is involved in but we don’t know what it does? Hint: 3 pipes; filter where bp!="biological process unknown" & mf=="molecular function unknown", and after selecting columns of interest, pipe the output to distinct(). The answer should be 737, and here are a few:
# A tibble: 737 × 3
symbol bp mf
<chr> <chr> <chr>
1 SFB2 ER to Golgi transport molec…
2 EDC3 deadenylylation-independent decapping molec…
3 PER1 response to unfolded protein* molec…
4 PEX25 peroxisome organization and biogenesis* molec…
5 BNI5 cytokinesis* molec…
6 CSN12 adaptation to pheromone during conjugation with cellular fusion molec…
7 SEC39 secretory pathway molec…
8 ABC1 ubiquinone biosynthesis molec…
9 PRP46 nuclear mRNA splicing, via spliceosome molec…
10 MAM3 mitochondrion organization and biogenesis* molec…
# ℹ 727 more rows
Exercise 9
When the growth rate is restricted to 0.05 by limiting Glucose, which biological processes are the most upregulated? Show a sorted list with the most upregulated BPs on top, displaying the biological process and the average expression of all genes in that process rounded to two digits. Hint: 5 pipes: filter, group_by, summarize, mutate, arrange.
# A tibble: 881 × 2
bp meanexp
<chr> <dbl>
1 fermentation* 6.28
2 glyoxylate cycle 5.28
3 oxygen and reactive oxygen species metabolism 5.04
4 fumarate transport* 5.03
5 acetyl-CoA biosynthesis* 4.32
6 gluconeogenesis 3.64
7 fatty acid beta-oxidation 3.57
8 lactate transport 3.48
9 carnitine metabolism 3.3
10 alcohol metabolism* 3.25
# ℹ 871 more rows
Exercise 10
Group the data by limiting nutrient (primarily) then by biological process. Get the average expression for all genes annotated with each process, separately for each limiting nutrient, where the growth rate is restricted to 0.05. Arrange the result to show the most upregulated processes on top. The initial result will look like the result below. Pipe this output to a View() statement. What’s going on? Why didn’t the arrange() work? Hint: 5 pipes: filter, group_by, summarize, arrange, View.
# A tibble: 5,257 × 3
# Groups: nutrient [6]
nutrient bp meanexp
<chr> <chr> <dbl>
1 Ammonia allantoate transport 6.64
2 Ammonia amino acid transport* 6.64
3 Phosphate glycerophosphodiester transport 6.64
4 Glucose fermentation* 6.28
5 Ammonia allantoin transport 5.56
6 Glucose glyoxylate cycle 5.28
7 Ammonia proline catabolism* 5.14
8 Ammonia urea transport 5.14
9 Glucose oxygen and reactive oxygen species metabolism 5.04
10 Glucose fumarate transport* 5.03
# ℹ 5,247 more rows
Exercise 11
Let’s try to further process that result to get only the top three most upregulated biolgocal processes for each limiting nutrient. Google search “dplyr first result within group.” You’ll need a filter(row_number()......) in there somewhere. Hint: 5 pipes: filter, group_by, summarize, arrange, filter(row_number().... Note: dplyr’s pipe syntax used to be %.% before it changed to |>. So when looking around, you might still see some people use the old syntax. Now if you try to use the old syntax, you’ll get a deprecation warning.
There’s a slight problem with the examples above. We’re getting the average expression of all the biological processes separately by each nutrient. But some of these biological processes only have a single gene in them! If we tried to do the same thing to get the correlation between rate and expression, the calculation would work, but we’d get a warning about a standard deviation being zero. The correlation coefficient value that results is NA, i.e., missing. While we’re summarizing the correlation between rate and expression, let’s also show the number of distinct genes within each grouping.
Warning: There was 1 warning in `summarize()`.
ℹ In argument: `r = cor(rate, expression)`.
ℹ In group 110: `nutrient = "Ammonia"` and `bp = "allantoate transport"`.
Caused by warning in `cor()`:
! the standard deviation is zero
Take the above code and continue to process the result to show only results where the process has at least 5 genes. Add a column corresponding to the absolute value of the correlation coefficient, and show for each nutrient the singular process with the highest correlation between rate and expression, regardless of direction. Hint: 4 more pipes: filter, mutate, arrange, and filter again with row_number()==1. Ignore the warning.
