XCMS_and_R_13Nov2012 page 1 of 11

Using XCMS and R to explore LC-MS data-sets

You will need R and XCMS installed on your PC. These are free, and can be obtained from
http://www.stats.bri.ac.uk/R/ for R, and http://www.bioconductor.org/ for XCMS.

(1) File format conversion

XCMS needs its files in an open-access common format, not in the proprietary format of a
manufacturer. For Thermo files I use ReadW which is available from
http://sourceforge.net/projects/sashimi/files/ to do the conversion, but there are others (e.g.
MSconvert). The best format for XCMS is mzXML, although it can use some others.

To convert, copy the readw.exe file into the same place as the data files (filename.raw) and
use the dos for instruction, which allows you to repeat an instruction for a large number of
files.
For %a in (*.raw) do readw %a

(2) Sort out the files into folders

XCMS was designed to compare two treatments, and expects the filename.mzXML files to
be distributed between two folders to match the treatments. Even if you have multiple
treatments it helps to make the two subfolders. If I have carried out QC runs using repeated
injections of a mixed sample, I put these in one folder, and all the sample files in the other.
Do not include blanks and standards (as these wont have peaks for XCMS to find and
match).

(3) Peak finding with XCMS

Open R, and tell it that we would like to use the XCMS library, which you should already
have installed (using the Packages menu). You only need to install it once, but every time
you open R you need to tell it that you will be using XCMS:
library(xcms)
Note that XCMS is case-sensitive.
Using the File menu, Change dir, navigate to the directory that contains your two sub-
directories of data.
Now create an object (you can use any name, e.g. A), and put in it the results of XCMSs
function xcmsSet, which makes a peak-list for each individual file included in the
subdirectories:
A <- xcmsSet()
You do not generally need to use any parameters with xcmsSet because the defaults are
good enough. If you use data from UPLC, CE or GC, with narrower peak-widths than
conventional LC, you will need to explore parameters. Typically you might have to specify
the peak-width at half-height (fwhm) in seconds. If you expect many peaks of the same mass
at different retention times (common in GC-MS, where many chemicals share fragments),
you need to increase the default maximum number of peaks per selected mass chromatogram:
B <- xcmsSet(fwhm=5, max=30)
A is a very large object containing lots of information. If you want to investigate it, you
might first like to tell R never to print out more than 50 lines of information about repetitive
objects:
options(max.print = 50)
XCMS_and_R_13Nov2012 page 2 of 11
If you cant remember how to use an instruction, or what parameters it can take, there is a
help-system.
help(options)
To find out more about A, or any other object you create, you can type:
attributes(A)
This tells you what sorts of things A contains, but not the actual values. If you want to see the
value held by an object, just type its name
A
In the case of a huge object like A, you wont see every value, just some predefined
information that the writers of xcmsSet thought might be helpful. Smaller, simpler objects
just print their values. For example you could make an object B and assign it the value 3:
B <- 3
This will return NULL if you look at its attributes, because it is so small that R hasnt
bothered remembering anything about it. If you type B, you will get the result 3.

A is a collection of peak-lists, but at the moment each data-file has its own peak-list, and they
are not cross-referenced. Peak 1 in file A is not necessarily the same as Peak 1 in file B. The
next step is to let XCMS create a grouping of peaks such that all peaks of similar mass and
retention time in all files are given a common number and recognised as a signal derived
from one single chemical.
A <- group(A)

Using peaks that are common to all samples, and reasonably intense, XCMS can estimate
how retention times have varied between runs. Using a running average of the marker
peaks that it has chosen, it will recalibrate the time-axis of all the samples. This is done by
retcor:
A2 <- retcor(A, family=s, plottype=m)
Note that we have used some parameters. Because R is case-sensitive, if you pass a parameter
with the wrong capitalisation, it wont be recognised. R is object-orientated, which means
each function is built on others, and parameters can be passed to function1 which are actually
only used by subfunction2; as a result, function1 will simply pass anything it doesnt
recognise to subfunction2 without complaint. In fact no function will complain about a
spurious parameter, so if you type Family instead of family you will get no error, but the
function wont do what you expect. Beware!
Some parameters need not be passed in full. s is short for symmetric, which means a re-
descending M estimator is used with Tukey's biweight function, allowing outlier removal
(this is obviously exactly what we intended; if you dont know what a redescending M
estimator is)


XCMS_and_R_13Nov2012 page 3 of 11
Retention time correction should produce a plot:

This shows the marker peaks as symbols, plotted at their corrected retention time on the x-
axis, and with their deviation in each sample in seconds as the y-axis. The lines are a moving
average that will be used to correct retention times of all remaining peaks, each sample
having its own moving average.

Sometimes if there are not many marker peaks, or they are unevenly spaced, there may be
wild digressions far beyond reasonable error. If this happens, you may prefer a linear best fit:
A2 <- retcor(A, family=s, plottype=m, method=linear)

We assign the results to a new object, A2, because we have actually changed the retention
times of the peaks in the peak-list.
-
6
0
-
4
0
-
2
0
0
2
0
Retention Time Deviation vs. Retention Time
Retention Time
R
e
t
e
n
t
i
o
n

T
i
m
e

D
e
v
i
a
t
i
o
n
CoraA
CoraB
CoraC
CoraD
CoraE
EA
EB
EC
ED
EE
EF
LBEA
LBEB
LBED
LBEE
0 500 1000 1500
Retention Time
P
e
a
k

D
e
n
s
i
t
y
All
Correction
-
5
0
-
4
0
-
3
0
-
2
0
-
1
0
0
1
0
Retention Time Deviation vs. Retention Time
Retention Time
R
e
t
e
n
t
i
o
n

T
i
m
e

D
e
v
i
a
t
i
o
n
CoraA
CoraB
CoraC
CoraD
CoraE
EA
EB
EC
ED
EE
EF
LBEA
LBEB
LBED
LBEE
0 500 1000 1500
Retention Time
P
e
a
k

D
e
n
s
i
t
y
All
Correction
XCMS_and_R_13Nov2012 page 4 of 11
If retention time correction fails, it is possible to go ahead using the un-corrected data.

A2 is no longer grouped, because weve moved all the peaks, and some may no longer be in
the right place to group with their original groupees, while others will have been moved
closer to their equivalents in other files, and can now be grouped. We re-group, but with
higher stringency:
A2 <- group(A2, bw=10)

Some peaks will not exist in some files. Rather than create a peak-list full of zero values,
which can cause problems to downstream statistics, we want a best estimate of the real signal
corresponding to that mass at that retention time in the samples where no peak was found.
XCMS adds up the baseline in this case:
A3 <- fillPeaks(A2)
Finally, we create the table of peak-areas and other information. The original intent of XCMS
was to look for differences, so this is a difference report.
Reporttab <- diffreport(A3, SubdirectoryA, SubdirectoryB, myOutput, 5)
SubdirectoryA/B are where the original mzXML files were placed. MyOutput will become a
tab-separated value file myOutput.tsv containing the peak-list. The number at the end is a
strange thing:

Diffreport expects to compare two treatments, so it will carry out t-tests for each peak in the
two subdirectory-groups, and it will order the results by decreasing significance. It expects
that we will want to look at the most significant peaks to check they are real. Therefore it
produces graphic files for the top n, which we specify. This is the 5. Since our t-test
comparison is meaningless if we have more than 2 treatments, we dont really need to
produce these graphics, but if we type 1 or less, we get a strange error. Hence 5.

Diffreport does create a new object, Reporttab, but I do not recommend you try to use it. I
find that columns of data within it that should be numerical become factors (non-numerical
data), and I havent found a way to make R treat them as numerical inside R. The best thing
is to re-import the tab-separated value file, which (usually) seems to turn numerical values
back into numbers.

(4) Trimming bad peaks

Some of the peaks that xcms has found are likely to be very small, or noise. Its possible to
carry on and analyse data without removing these peaks, but they will make it harder to find
the true differences. The best approach is to use a QC sample, a mixture of all the other
samples (therefore containing all peaks) run repeatedly throughout the sequence. This should
give the same area for every peak each time; peaks which vary are probably bad peaks. I
show below how to do this in R. If a QC sample is not available, then we can look at the
variability within replicates of a sample; I show this for Excel.

(4a) Using a QC sample

Read in the diffreport output into some convenient object:
Rtb <- read.table(myOutput.tsv)
Check what columns it contains, and find their numbers:
attributes(Rtb)
XCMS_and_R_13Nov2012 page 5 of 11
Rtb is a list of columns, each of which is itself a list of something else. R allows us to refer to
just some of the columns. If our QC samples have their data in columns 14 to 17, we can
select just these and put them in a separate table:
QCtb <- Rtb[14:17]
The measure of variability is the relative standard deviation. For this we first need the actual
standard deviation of each row. Unfortunately Rs instruction for calculating standard
deviations does it for each column. Fortunately, we can transpose our table (turn rows into
columns. But the transpose instruction (t) expects the data to be a matrix (2-dimensional grid
of numbers) whereas what we actually have is a thing called a data-frame (a list of vectors; a
list of columns that are all the same length, but not necessarily all containing the same sort of
data). Fortunately since our data-frame has columns all of which are numerical, we can turn it
into a matrix using as.matrix. These are all combined:
QCsd <- sd(t(as.matrix(QCtb)))
This produces a vector of numbers (list of numbers) containing a standard deviation for each
peak. Now we need the means, which fortunately can be calculated directly from rows:
QCmean <- RowMeans(QCtb)
Finally we divide standard deviations by means to get the relative standard deviations:
RSD <- QCsd / QCmean
Note that this does them all in one go. This is the advantage of R.

The point of all this is that we now want to select peaks (rows in our original table) where the
RSD is less than some suitable value (e.g. 30% = 0.3). R refers to anything in a matrix by its
coordinates, [row, column]. Really cleverly, we can refer to rows by number, or by a
condition:
RtbTrimmed <- Rtb[RSD < 0.3,]
What is really happening here is that RSD < 0.3 is creating a vector (list) of TRUE/FALSE
values, and Rtb[###,] is selecting all rows where the vector contains the value TRUE. The
comma is telling it that were referring to rows, and we want all columns.

RtbTrimmed is now a shorter table containing only good peaks, and we can save it or use it
for PCA. To save it:
write.csv(RtbTrimmed, TrimmedOutput.csv)

Incidentally, you can also see the RSD values as a histogram or a plot, and view what sort of
cut you are making:
hist(RSD)
plot(RSD)

XCMS_and_R_13Nov2012 page 6 of 11

(4b) Using Excel with no QC sample

In Excel, calculate the relative standard deviation of each treatment in a separate column
using expressions such as:
=STDEV(T2:Y2)/AVERAGE(T2:Y2)
Now find the minimum RSD. This makes sense because a peak might be very clearly
measured in treatment A, but completely absent from treatment B. In this case it will have a
good RSD for A, but a terrible one in B. Make a new column for the minima:
=MIN(AD2:AF2)
Sort the table in order of increasing RSD, and take only the rows you want. Save the trimmed
table as a comma-separated or tab-separated value file (R cannot read Excel workbooks).

(5) Multiple-univariate statistics

We now need to look which peaks have changed. It is possible to do multiple t-tests in Excel.
The problem with this is that the more t-tests we do, the greater our risk of false positives
(expect 1 per 100 rows at P<0.01). Classically the solution is Bonferroni correction, which is
simply altering the P cut-off to 0.01/100 if working with 100 independent rows. This is over-
cautious since many of our rows will be different ions derived from the same chemical, so we
dont have 100 independent rows. It doesnt work for us. Multiple t-tests will cause you
trouble with statistically-aware reviewers.

The next possibility is multiple Anova. I still feel this suffers from the same problem, but
reviewers dont seem to distrust it as badly, and it is more appropriate where we have not
only multiple peaks, but multiple (more than 2) treatments. Its hard to do in Excel because
the Anova feature of Excel is part of the Analysis tool-pack, and returns a block of data, not a
single number (and has to be used via a wizard).

Anova expects there to be equal numbers of replicates of each treatment. If you dont have
this, the JIC statistics courses offer alternatives.

It may be possible to do multiple anova in Genstat. An alternative is to use R.

This is a little clumsy, but I provide a method below. Im not sure it should be necessary to
make a new function to carry out multiple Anova, but I did.
Histogram of qcrsd
qcrsd
F
r
e
q
u
e
n
c
y
0.0 0.5 1.0 1.5 2.0
0
1
0
2
0
3
0
4
0
5
0
6
0
0 50 100 150
0
.
0
0
.
5
1
.
0
1
.
5
Index
q
c
r
s
d
XCMS_and_R_13Nov2012 page 7 of 11

First create a function to do the job:
MultipleAnova <- function(indata, inlevels) {}
This is currently empty. Edit it, putting the edited version back into the original object (the
original function).
MultipleAnova <- edit(MultipleAnova)
This will open a text editor, in which you can put the following code:
function(indata, inlevels)
## carries out repeated one-way anova
{
Pvector <- c()
indata <- as.matrix(indata)
qrows <- attr(indata, "dim")[1]
for (q in 1:qrows)
{
subdata <- indata[q,]
submodel <- lm(subdata~inlevels)
subanova <- anova(submodel)
Pvector <- c(Pvector, subanova$"Pr(>F)"[1])
}
Pvector
}

Refer to section 4a on how to read in your peak-areas, and select from the whole table a sub-
table containing the columns you want. Create a sub-table containing just the peak-areas from
the samples to be analysed. Your actions might look a bit like this:
dta <- read.csv("SLloyd_1841_3Nov2012_Export.csv")
options(max.print = 50)
attr(dta)
attributes(dta)
adta <- dta[,6:17]
attributes(adta)
Note that we checked what columns we had using attributes.

Now we create a model, a description of our experiment, for R. We should have grouped the
replicates of the treatment together. For example, if we have three treatments, 4 replicates
each, wed create a model dtalevels using gl:
dtalevels <- gl(3, 4, 12, labels=c("A", "B", "C"))
dtalevels
The second line just checks what the model looks like. Now we run the multiple anovas:
tempPvec <- multipleanova(adta, dtalevels)
write.csv(tempPvec, "multanova.csv")
The results can be combined with the full peak-list in Excel, taking care that we dont
combine two peak-lists sorted differently!


XCMS_and_R_13Nov2012 page 8 of 11
(6) Multivariate statistics

There are many different approaches, the commonest being PCA and PLS-DA. PCA is very
user-friendly and safe; if you wish to use PLS-DA, its a good idea to discuss it with experts
(e.g. Kate Kemsley and her group at IFR).

Both PLS-DA and PCA are based on the idea of reducing a 300-dimensional data-set to
something we can see.

Introduction to PCA

Please skip this section if you are already familiar with the method.

PCA is a traditional approach to looking at the differences between two (or more) sets of
samples. The idea is that each sample gives a set of measurements, each corresponding to a
different chemical. If a sample only had two chemicals, we could plot it in 2-dimensional
space, each chemical being one axis of our plot, and each sample giving one dot on the plot.
In real examples we might have hundreds, even thousands of values (though probably many
less chemicals, as one chemical will create multiple ions). We therefore need a plot with
more than 3 dimensions (in fact hundreds or thousands), which cannot be imagined. PCA
provides a way to reduce this plot to two dimensions, on paper.

The idea is that the experiment will produce a cloud of spots. Hopefully the spots arent
scattered randomly, but treated samples will contain different metabolites than controls.
Treated samples will be found on one side of the cloud while the controls are the other. PCA
merely turns the axes round so that one of the axes points along the direction in which the
cloud is most stretched. Ideally the treated samples are a long way from the controls, so the
cloud is longest in that direction! In this case, PCA finds the most obvious differences
between controls and treated samples. It progresses to put the second axis at right-angles to
the first, but in the second-most-stretched direction in the cloud, and so on.

The diagram below shows two treatments, red and blue, with just two chemicals. In this
experiment there is some other cause of scatter, and the cloud is long in the wrong
direction. PCA will draw its first axis in this direction. We could re-plot the data as points
moved onto this axis, reducing the number of dimensions, but it wouldnt tell us what we
want to know:
XCMS_and_R_13Nov2012 page 9 of 11

On the other hand, the second axis at right-angles to the first, does contain the right
information:


In the same way that we can move points sideways from the two-dimensional plot and create
a one-dimensional plot in which the treatments are separated, PCA can be used to reduce a
very large number of dimensions (100s to 1000s) to just two, which is can be drawn on
paper.

Note that PCA rotates the axes through the origin, which means the points should be scattered
around the origin. For this reason the data should be centred, i.e. the mean is subtracted from
the data. Some metabolites are more abundant than others, so the cloud of points will be more
stretched along their axes. For this reason, PCA is biased towards abundant chemicals. This
can be avoided by scaling each chemical (dividing all the measurements of each chemical by
the standard deviation of the measurements for that chemical).

PCA is closely related to partial least squares regression (PLS) in which the axes still exist,
but are rotated so that they distinguish most clearly the different treatments. This is a very
sensitive method, but certain to produce false positives. It is only safe when used in
Chemical A
-2 -1 0 1 2
C
h
e
m
i
c
a
l

B
-2
-1
0
1
2
Chemical A
-2 -1 0 1 2
C
h
e
m
i
c
a
l

B
-2
-1
0
1
2
XCMS_and_R_13Nov2012 page 10 of 11
conjunction with carefully planned controls (typically by rotating the axes using some data,
and testing the validity of the result with other data, for which reason PLS is even more
dependent than PCA on having lots of replicate samples).

Doing it!

First import the data from the trimmed peak-list in Excel (or R directly if you trimmed using
R and a QC sample, see section 4a).
Rtab <- read.csv(TrimmedData.csv)
Using attributes, find the columns that contain sample data:
attributes(Rtab)
samptab <- Rtab[,15:38]
This selects, as an example, all rows (before comma) from columns 15-38 (after comma).
The R function that does PCA expects the data in rows rather than columns, so we need to
transpose:
samptabT <- t(samptab)
Create a list of labels corresponding to the treatments of each column. These should be in the
right order, and should be single letters as they will be used for plotting. Lists can be created
by concatenating, using c:
sampLabels <- c(A, W, b, A, A, b, W, W, b)
Now carry out the principle component analysis, putting the results in a new object:
pca <- prcomp(samptabT, center=TRUE, scale=TRUE, xret=TRUE)
This automatically centres the peak areas on zero (subtracts the mean) and scales them such
that they all have a standard deviation of 1. If you dont scale, you will probably find that the
greatest variability (PC1) is always the most abundant chemical, even if it doesnt tell you
anything about your experiment. Xret ensures that the actual data are put in the object so we
can plot our PCA plots.

To plot PC1 versus PC2,:
Plot(pca$x[,1], pca$x[,2], pch=sampLabels)

XCMS_and_R_13Nov2012 page 11 of 11

Having seen that PC2 in this case explains the difference between L and C, while PC1
contains the difference between L and E, we want to know which peaks contribute to
each PC. The contributions are in the loadings, which R refers to as rotations. We can export
these, for the first three PCs:
Write.csv(pca$rotation*,1:3+, rotation_values.csv)

We would then attach them to the Excel peak-list, and add a column containing the absolute
values (i.e. ignoring the sign). The sign of the rotation tells you whether a chemical went
down or up on treatment, and is arbitrary (different software packages will reverse the axis).
Big negative values are just as important as big positive ones. In Excel:
=ABS(AH2)
Sort by this column (largest to smallest) and then explore what the peaks are. You should
expect to find multiple features (rows) at nearly the same retention time, meaning that an ion
has changed, and so has its isotope peak, and some related fragments

C
C
C
C
C
E
E
E
E
E
E
L L
L
L
-10 0 10 20 30
-
1
5
-
1
0
-
5
0
5
1
0
pca$x[, 1]
p
c
a
$
x
[
,

2
]