Intro to RaMS

TICs, BPCs, and metadata

A TIC reports the total intensity measured by the mass analyzer during each scan, so the data is parsed into two columns: retention time (rt) and intensity (int). This makes it easy to read and simple to plot:

single_file <- data_files[2]

msdata <- grabMSdata(single_file, grab_what = "TIC")
#> 
#> Reading file LB12HL_AB.mzML.gz... 0.06 secs 
#> Reading TIC...0.03 secs 
#> Total time: 0.09 secs

knitr::kable(head(msdata$TIC, 3))

Since we asked for a single thing, the TIC, our file_data object is a list with a single entry: the TIC. Let’s plot that data:

plot(msdata$TIC$rt, msdata$TIC$int, type = "l")

Simple enough!

A BPC is just like a TIC except that it records the maximum intensity measured, rather than the sum of all intensities. This data is also collected by the mass analyzer and doesn’t need to be calculated.

msdata <- grabMSdata(single_file, grab_what = "BPC")
#> 
#> Reading file LB12HL_AB.mzML.gz... 0.07 secs 
#> Reading BPC...0.03 secs 
#> Total time: 0.1 secs

Since the data is parsed in a “tidy” format, it plays nicely with popular packages such as ggplot2. Let’s use that to plot our BPC instead of the base R plotting system:

library(tidyverse)
ggplot(msdata$BPC) + geom_line(aes(x=rt, y=int))

The advantages of tidy data and ggplot become clear when we load more than one file at a time because we can group and color by the third column, the name of the file from which the data was read. RaMS will also automatically enable a loading bar if more than one file is loaded; this can be disabled by setting the verbosity parameter to 0.

msdata <- grabMSdata(data_files[2:4], grab_what = "BPC")
#> 
  |                                                                            
  |                                                                      |   0%
  |                                                                            
  |=======================                                               |  33%
  |                                                                            
  |===============================================                       |  67%
  |                                                                            
  |======================================================================| 100%
#> Total time: 0.26 secs

ggplot(msdata$BPC) + geom_line(aes(x=rt, y=int, color=filename))

And of course, this means that all of ggplot’s aesthetic power can be brought to your chromatograms as well, so customize away!

ggplot(msdata$BPC) + 
  geom_polygon(aes(x=rt, y=int, color=filename), lwd=1, fill="#FFFFFF44") +
  theme(legend.position = c(0.8, 0.7), plot.title = element_text(face="bold"),
        axis.title = element_text(size = 15)) +
  scale_y_continuous(labels = c(0, "250M", "500M"), breaks = c(0, 2.5e8, 5e8)) +
  scale_colour_manual(values = c("#2596be", "#6c25be", "#bea925")) +
  labs(x="Retention time (minutes)", y="Intensity", 
       title = "Base peak chromatogram", color="Files:") +
  coord_cartesian(xlim = c(7.50, 9), ylim = c(0, 5e8))

RaMS also provides some basic file metadata extraction capability, although the focus for this package is on the actual data and other MS packages handle file metadata much more elegantly. This is one area where there are major differences between mzML and mzXML file types - the mzXML file type simply doesn’t encode as much metadata as the mzML filetype, so RaMS can’t extract it.

# Since the minification process strips some metadata, I use the 
# less-minified DDA file here
grabMSdata(files = data_files[1], grab_what = "metadata")
#> 
#> Reading file DDApos_2.mzML.gz... 0.58 secs 
#> Reading file metadata...0.18 secs 
#> Total time: 0.77 secs
#> $metadata
#>                                   source_file  inst_data       config_data
#> 1: 170223_Poo_AllCyanoAqExtracts_DDApos_2.raw None found <data.frame[1x3]>
#>    timestamp mslevels polarity         filename
#> 1:      <NA> MS1, MS2 positive DDApos_2.mzML.gz

Adding a column: MS1 data

MS1 data can be extracted just as easily, by supplying “MS1” to the grab_what argument of grabMSdata function.

msdata <- grabMSdata(data_files[2:4], grab_what = "MS1")
#> 
  |                                                                            
  |                                                                      |   0%
  |                                                                            
  |=======================                                               |  33%
  |                                                                            
  |===============================================                       |  67%
  |                                                                            
  |======================================================================| 100%
#> Total time: 0.31 secs

knitr::kable(head(msdata$MS1, 3))

So we’ve now got the mz column, corresponding to the mass-to-charge ratio (m/z) of an ion. This means that we can now filter our data for specific masses and separate out molecules with different masses.

Note that this also makes the data much larger in R’s memory - so don’t go loading hundreds of files simultaneously. If that’s necessary, check out the section below on saving space.

Because RaMS returns data.tables rather than normal data.frames, indexing is super-fast and a bit more intuitive than with base R. Below, I also use the pmppm function from RaMS to produce a mass range from an initial mass and spectrometer accuracy (here, 5 parts-per-million).

library(data.table)

adenine_mz <- 136.06232

adenine_data <- msdata$MS1[mz%between%pmppm(adenine_mz, ppm=5)]

ggplot(adenine_data) + geom_line(aes(x=rt, y=int, color=filename))

This makes it easy to grab the data for multiple compounds of interest with a simple loop, provided here by the purrr package of the tidyverse:

mzs_of_interest <- c(adenine=136.06232, valine=118.0865, homarine=138.055503)

mass_data <- imap_dfr(mzs_of_interest, function(mz_i, name){
  cbind(msdata$MS1[mz%between%pmppm(mz_i, ppm=10)], name)
})

ggplot(mass_data) + 
  geom_line(aes(x=rt, y=int, color=filename)) + 
  facet_wrap(~name, ncol = 1, scales = "free_y")

Moving along: MS2 data

RaMS also handles MS² data elegantly. Request it with the “MS2” option for grab_what, although it’s often a good idea to grab the MS1 data alongside.

msdata <- grabMSdata(data_files[1], grab_what = "everything")
#> 
#> Reading file DDApos_2.mzML.gz... 0.56 secs 
#> Reading MS1 data...0.18 secs 
#> Reading MS2 data...0.48 secs 
#> Reading BPC...0.11 secs 
#> Reading TIC...0.11 secs 
#> Reading file metadata...0.19 secs 
#> Total time: 1.62 secs
knitr::kable(head(msdata$MS2, 3))

DDA data can be plotted nicely with ggplot2 as well. Typically it makes sense to filter for a precursor mass, then render the fragments obtained.

homarine_mz <- 137.047678+1.007276

homarine_MS2 <- msdata$MS2[premz%between%pmppm(homarine_mz, 5)]
homarine_MS2$int <- homarine_MS2$int/max(homarine_MS2$int)

ggplot(homarine_MS2) +
  geom_point(aes(x=fragmz, y=int)) +
  geom_segment(aes(x=fragmz, xend=fragmz, y=int, yend=0)) +
  scale_y_continuous(breaks = c(0, 50, 100), labels = c("0%", "50%", "100%"))

  labs(x="Fragment m/z", y="")
#> $x
#> [1] "Fragment m/z"
#> 
#> $y
#> [1] ""
#> 
#> attr(,"class")
#> [1] "labels"

This is also the perfect place to enable some interactivity with packages such as plotly, making data exploration not only simple but also enjoyable.

## Not run to save space in the vignette:
library(plotly)
msdata <- grabMSdata(data_files, grab_what = c("MS1", "MS2", "BPC"))

compound_MS1 <- msdata$MS1 %>% 
  filter(mz%between%pmppm(homarine_mz, 10)) %>%
  filter(!str_detect(filename, "DDA"))

compound_MS2 <- msdata$MS2[premz%between%pmppm(homarine_mz, 10)] %>% 
  group_by(rt) %>%
  arrange(desc(int)) %>%
  summarise(frags=paste(
    paste(round(fragmz, digits = 3), round(int), sep = ": "), collapse = "\n"),
    .groups="drop"
  ) %>%
  mutate(int=approx(x = compound_MS1$rt, y=compound_MS1$int, xout = rt)$y)
plot_ly(compound_MS1) %>%
  add_trace(type="scatter", mode="lines", x=~rt, y=~int, color=~filename,
            hoverinfo="none") %>%
  add_trace(type="scatter", mode="markers", x=~rt, y=~int,
            text=~frags, hoverinfo="text", showlegend=FALSE,
            marker=list(color="black"), data = compound_MS2) %>%
  layout(annotations=list(x=min(compound_MS2$rt), y=median(compound_MS2$int)*10, 
                          text="Mouse over to see\nMSMS fragments"))

Easy access to MS² data also allows us to rapidly perform simple operations such as searching for a specific fragment mass. For example, if we know that homarine typically produces a fragment with a mass of 94.06567, we simply subset the MS² data for fragments in a range around that mass:

homarine_frag_mz <- 94.06567
msdata$MS2[fragmz%between%pmppm(homarine_frag_mz, ppm = 5)] %>%
  head() %>% knitr::kable()

We find that there’s not only homarine that produces that fragment, but several another compounds. Precursor masses like this can then be searched manually in online databases or, since the data is already in R, passed to a script that automatically searches them.

Similarly, we can easily search instead for neutral losses with this method. If we suspect other molecules are producing a similar neutral loss as homarine:

homarine_neutral_loss <- homarine_mz - homarine_frag_mz

msdata$MS2 <- mutate(msdata$MS2, neutral_loss=premz-fragmz) %>%
  select("rt", "premz", "fragmz", "neutral_loss", "int", "voltage", "filename")
msdata$MS2[neutral_loss%between%pmppm(homarine_neutral_loss, ppm = 5)] %>%
  arrange(desc(int)) %>% head() %>% knitr::kable()

We can again confirm our suspicions that there’s another signal with a similar neutral loss: one with a mass of 137.0470.

Saving space: EICs and rtrange

Mass-spec files are typically tens or hundreds of megabytes in size, which means that the simultaneous analysis of many files can easily exceed a computer’s memory. Since RaMS stores all data in R’s working memory, this can become a problem for large analyses.

However, much of the usage envisioned for RaMS on this scale doesn’t require access to the entire file, the entire time. Instead, users are typically interested in a few masses of interest or a specific time window. This means that while each file still needs to be read into R in full to find the data of interest, extraneous data can be discarded before the next file is loaded.

This functionality can be enabled by passing “EIC” and/or “EIC_MS2” to the grab_what argument of grabMSdata, along with a vector of masses to extract (mz) and the instrument’s ppm accuracy. When this is enabled, files are read into R’s memory sequentially, the mass window is extracted, and the rest of the data is discarded.

all_data <- grabMSdata(data_files, grab_what = c("MS1", "MS2"))
#> 
  |                                                                            
  |                                                                      |   0%
  |                                                                            
  |==================                                                    |  25%
  |                                                                            
  |===================================                                   |  50%
  |                                                                            
  |====================================================                  |  75%
  |                                                                            
  |======================================================================| 100%
#> Total time: 1.57 secs

mzs_of_interest <- c(adenine=136.06232, valine=118.0865)
small_data <- grabMSdata(data_files, grab_what = c("EIC", "EIC_MS2"),
                         mz=mzs_of_interest, ppm = 5)
#> 
  |                                                                            
  |                                                                      |   0%
  |                                                                            
  |==================                                                    |  25%
  |                                                                            
  |===================================                                   |  50%
  |                                                                            
  |====================================================                  |  75%
  |                                                                            
  |======================================================================| 100%
#> Total time: 1.79 secs

all_data$MS1 %>%
  mutate(type="All data") %>%
  rbind(small_data$EIC %>% mutate(type="Extracted data only")) %>%
  filter(!str_detect(filename, "DDA")) %>%
  group_by(rt, filename, type) %>%
  summarise(TIC=sum(int), .groups="drop") %>%
  ggplot() +
  geom_line(aes(x=rt, y=TIC, color=filename)) +
  facet_wrap(~type, ncol = 1) +
  theme(legend.position = "top")

# Size reduction factor:
as.numeric(object.size(all_data)/object.size(small_data))
#> [1] 50.70094

As expected, the size of the small_data object is much smaller than the all_data object, here by a factor of about 50x. For files that haven’t already been “minified”, that size reduction will be even more significant. Of course, this comes with the cost of needing to re-load the data a second time if a new mass feature becomes of interest but this shrinkage is expecially valuable for targeted analyses where the analytes of interest are known in advance.

A second way of reducing file size is to constrain the retention time dimension rather than the m/z dimension. This can be done with the rtrange argument, which expects a length-two vector corresponding to the upper and lower bounds on retention time. This is useful when a small time window is of interest, and only the data between those bounds is relevant.

small_data <- grabMSdata(data_files, grab_what = c("MS1", "MS2"), rtrange = c(5.5, 7))
#> 
  |                                                                            
  |                                                                      |   0%
  |                                                                            
  |==================                                                    |  25%
  |                                                                            
  |===================================                                   |  50%
  |                                                                            
  |====================================================                  |  75%
  |                                                                            
  |======================================================================| 100%
#> Total time: 1.2 secs

all_data$MS1 %>%
  mutate(type="All data") %>%
  rbind(small_data$MS1 %>% mutate(type="Extracted data only")) %>%
  filter(!str_detect(filename, "DDA")) %>%
  group_by(rt, filename, type) %>%
  summarise(TIC=sum(int), .groups="drop") %>%
  ggplot() +
  geom_line(aes(x=rt, y=TIC, color=filename)) +
  facet_wrap(~type, ncol = 1) +
  theme(legend.position = "top")

# Size reduction factor:
as.numeric(object.size(all_data)/object.size(small_data))
#> [1] 3.90223

The savings are more modest here, but constraining the retention time actually speeds up data retrieval slightly. Since m/z and intensity data is encoded in mzML and mzXML files while retention time information is not, eliminating scans with a retention time window removes the need to decode the intensity and m/z information in those scans.

However, decoding is rarely the rate-limiting step and for more information about speeding things up, continue to the next section.

Speeding things up

So, RaMS isn’t fast enough for you? Let’s see what we can do to improve that. The first step in speeding things up is discovering what’s slow. Typically, this is the process of reading in the mzML/mzXML file rather than any processing that occurs afterward, but this is not always the case. To examine bottlenecks, RaMS includes timing information that’s produced if the verbosity argument is set to 2 or higher.

all_data <- grabMSdata(data_files, grab_what = c("MS1", "MS2"), verbosity = 2)
#> 
  |                                                                            
  |                                                                      |   0%
#> Reading file DDApos_2.mzML.gz... 0.54 secs 
#> Reading MS1 data...0.19 secs 
#> Reading MS2 data...0.51 secs 
#> 
  |                                                                            
  |==================                                                    |  25%
#> Reading file LB12HL_AB.mzML.gz... 0.04 secs 
#> Reading MS1 data...0.05 secs 
#> Reading MS2 data...0 secs 
#> 
  |                                                                            
  |===================================                                   |  50%
#> Reading file LB12HL_CD.mzML.gz... 0.06 secs 
#> Reading MS1 data...0.05 secs 
#> Reading MS2 data...0.02 secs 
#> 
  |                                                                            
  |====================================================                  |  75%
#> Reading file LB12HL_EF.mzML.gz... 0.05 secs 
#> Reading MS1 data...0.05 secs 
#> Reading MS2 data...0 secs 
#> 
  |                                                                            
  |======================================================================| 100%
#> Total time: 1.54 secs

In general, slow file read times can be improved by compressing the data. mzML files are highly compressible, with options to compress both the data itself using the zlib method and the files as a whole using gzip.

gzip is the simplest one to use, as a plethora of methods exist to compress files this way, including online sites. RaMS can read the data directly from a gzipped file, no decompression necessary, so this is an easy way to reduce file size and read times. Sharp-eyed users will have noticed that the demo files are already gzipped, which is part of the reason they are so small.

zlib compression is slightly trickier, and is most often performed with tools such as Proteowizard’s msconvert tool with the option “-z”.

Read times may also be slow if files are being accessed via the Internet, either through a VPN or network drive. If your files are stored elsewhere, consider first moving those files somewhere more local before reading data from them.

If the bottleneck appears when reading MS¹ data, consider restricting the retention time range with the rtrange argument or using more detailed profiling tools such as RStudio’s “Profile” options or the profvis package. Pull requests that improve data processing speed are always welcome!

While the package is not currently set up for parallel processing, this is a potential future feature if a strong need is demonstrated.

Finer control: grabMzmlData, grabMzxmlData

While there’s only one main function (grabMSdata) in RaMS, you may have noticed that two other functions have been exposed that perform similar tasks: grabMzmlData and grabMzxmlData. The main function grabMSdata serves as a wrapper around these two functions, which detects the file type, adds the “filename” column to the data, and loops over multiple files if provided. However, there’s often reason to use these internal functions separately.

For one, the objects themselves are smaller because they don’t have the filename column attached yet. You as the user will need to keep track of which data belongs to which files in this case.

Another use case might be applying functions to each file individually, perhaps aligning to a reference chromatogram or identifying peaks. Rather than spending the time to bind the files together and immediately separate them again, these functions have been exposed to skip that step.

Finally, these functions are useful for parallelization. Because iterating over each mass-spec file is often the largest reasonable chunk, these functions can be passed directly to parallel processes like mclapply or doParallel. However, parallelization is a beast best handled by individual users because its actual implementation often differs wildly and its utility depends strongly on individual setups (remember that parallelization won’t help with slow I/O times, so it may not always improve data processing speed.)

## Not run:
library(parallel)
cl <- makeCluster(getOption("cl.cores", detectCores()-1))
output_data <- parLapply(data_files, grabMzmlData, grab_what="everything", cl = cl)

library(foreach)
library(doParallel)
registerDoParallel(detectCores()-1)
output_data <- foreach (i=data_files) %dopar% {
  RaMS::grabMzmlData(i, grab_what="everything")
}
stopImplicitCluster()

The nitty-gritty details

RaMS is possible because mzML and mzXML documents are fundamentally XML-based. This means that we can leverage speedy and robust XML parsing packages such as xml2 to extract the data. Fundamentally, RaMS relies on XPath navigation to collect various bits of mass-spec data, and the format of mzML and mzXML files provides the tags necessary. That means a lot of RaMS code consists of lines like the following:

## Not run:
data_nodes <- xml2::xml_find_all(mzML_nodes, xpath="//d1:precursorMz")
raw_data <- xml2::xml_attr(data_nodes, "value")

…plus a lot of data handling to get the output into the tidy format.

The other tricky bit of data extraction is converting the (possibly compressed) binary data into R-friendly objects. This is usually handled with code like that shown below. Many of the settings can be deduced from the file, but sometimes compression types need to be guessed at and will throw a warning if so.

## Not run:
decoded_binary <- base64enc::base64decode(binary)
raw_binary <- as.raw(decoded_binary)
decomp_binary <- memDecompress(raw_binary, type = file_metadata$compression)
final_binary <- readBin(decomp_binary, what = "double",
                        n=length(decomp_binary)/file_metadata$mz_precision,
                        size = file_metadata$mz_precision)

# See https://github.com/ProteoWizard/pwiz/issues/1301

Fundamentally, mass-spectrometry data is formatted as a ragged array, with an unknown number of m/z and intensity values for a given scan. This makes it difficult to encode neatly without interpolating, but tidy data provides a solution by stacking those arrays rather than aligning them into some sort of matrix.

This ragged shape is also the reason that subsetting mass-spec data by retention time is trivial - grab the scans that correspond to the desired retention times and you’re done. Subsetting by mass, on the other hand, requires decoding each and every scan’s m/z and intensity data. If you’re reading a book and only want a couple chapters, it’s easy to flip to those sections. If you’re looking instead for every time a specific word shows up, you’ve gotta read the whole thing.

For more information about the mzML data format and its history, check out the specification at https://www.psidev.info/mzML.

Minified file details

While many RaMS users will arrive at this vignette with their own data, several small files are provided for demonstration. These files originate from the Ingalls Lab at the University of Washington, USA and are published in the manuscript “Metabolic consequences of cobalamin scarcity in diatoms as revealed through metabolomics”. Files were downloaded from the corresponding Metabolights repository.

Due to CRAN’s restrictions on package size, the original files cannot be included. Instead, MS1 data was compressed by extracting EICs for each compound in the Ingalls Lab list of targeted compounds as of August 24, 2020. EIC width was set with a ppm value of 10 and time data were filtered to fall between 4 and 9 minutes. These compressed files were then written out using MSnbase’s writeMSdata function after coercing the existing data frame into an MSnExp object. Finally, they were gzip-compressed with R.utils’s gzip function and default parameters.

MS2 data was processed via Proteowizard’s msconvert tool instead. Files were converted with the flags -gz and --noindex to save space, and three filters were used: - --filter "scanTime [240,540]" - --filter "mzWindow [0,120]" - --filter "absolute 100000 most-intense"

These filters reduced the MS2 data to a sufficiently limited size that they could be published with the package. The mzXML file was created identically, but the --mzXML flag was added.