# Combining HDMSe/MSe data using *synapter* to optimise identification and quantitation In synapter: Label-free data analysis pipeline for optimal identification and quantitation

suppressPackageStartupMessages(library("synapter"))
suppressPackageStartupMessages(library("synapterdata"))
suppressPackageStartupMessages(library("BiocStyle"))
suppressPackageStartupMessages(library("qvalue"))
## preparing some data
data(ups25b)  ## from synapterdata
res <- ups25b ## for synergise and plotRt






# Introduction{#sec:intro}

## Background{#sec:bkg}

The main functionality of r Biocpkg("synapter") is to combine proteomics data acquired under different mass spectrometry settings or with different samples to (i) optimise the respective qualities of the two data sets or (ii) increase the number of identifications, thereby decreasing missing values. Besides r Biocpkg("synapter") offers other functionality inaccessible in the default pipeline, like peptide FDR estimation and filtering on peptide match type and peptide uniqueness.

The example that motivated the development of this package was to combine data obtained on a Synapt G2 instrument:

1. HDMS$^E$ data, acquired with additional peptide separation using an ion mobility cell, thus leading to better (both in number and in quality) identification and
2. standard MS$^E$ data (acquired without ion mobility separation), providing better data quantitation.

The former is data is called identification peptides and the latter quantitation peptides, irrespective of the acquisition mode (HDMS$^E$ or MS$^E$). This HDMS$^E$/MS$^E$ design is used in this document to illustrate the r Biocpkg("synapter") package.

However, although HDMS$^E$ mode possesses superior identification and MS$^E$ mode superior quantitation capabilities and transferring identifications from HDMS$^E$ to MS$^E$ is a priori the most efficient setup, identifications can be transferred between any runs, independently of the acquisition mode. This allows to reduce the number of missing values, one of the primary limitation of label-free proteomics. Thus users will benefit from r Biocpkg("synapter")'s functionality even if they run their instruments in a single mode (HDMS$^E$ or MS$^E$ only).

However, as will be shown in section Data analysis, transferring identifications from multiple runs to each other increases analysis time and peptide FDR within the analysis. r Biocpkg("synapter") allows to minimise these effects to acceptable degree by choosing runs to transfer identifications from and merging them in the master HDMS$^E$ file.

This data processing methodology is described in section HDMS$^E$/MS$^E$ data analysis and the analysis pipeline is described in section Different pipelines.

To maximise the benefit of combining better identification and quantitation data, it is also possible to combine several, previously merged identification data files into one master set. This functionality is described in section Using master peptide files.

Finally, section Analysis of complex experiments illustrates a complete pipeline including r Biocpkg("synapter") and r Biocpkg("MSnbase") [@Gatto2012] packages to perform protein label-free quantitation: how to combine multiple r Biocpkg("synapter") results to represent the complete experimental design under study and further explore the data, normalise it and perform robust statistical data analysis inside the R environment.

The rationale underlying r Biocpkg("synapter")'s functionality are described in [@Shliaha2013] and [@Bond2013]. The first reference describes the benefits of ion mobility separation on identification and the effects on quantitation, that led to the development of r Biocpkg("synapter"), which in described and demonstrated in [@Bond2013].

r Biocpkg("synapter") is written for R~[@R], an open source, cross platform, freely available statistical computing environment and programming language^[https://www.r-project.org/]. Functionality available in the R environment can be extended though the usage of packages. Thousands of developers have contributed packages that are distributed via the Comprehensive R Archive Network (CRAN) or through specific initiatives like the Bioconductor^[https://www.bioconductor.org/] project [@Gentleman2004], focusing on the analysis and comprehension of high-throughput biological data.

r Biocpkg("synapter") is such an R package dedicated to the analysis of label-free proteomics data. To obtain detailed information about any function in the package, it is possible to access it's documentation by preceding it's name with a question mark at the command line prompt. For example, to obtain information about the r Biocpkg("synapter") package, one would type ?synapter.

## Installation{#sec:install}

r Biocpkg("synapter") is available through the Bioconductor project. Details about the package and the installation procedure can be found on its page^[https://bioconductor.org/packages/synapter/]. Briefly, installation of the package and all its dependencies should be done using the dedicated Bioconductor infrastructure as shown below:

source("https://bioconductor.org/biocLite.R")
## or, if you have already used the above before
library("BiocInstaller")
## and to install the package
biocLite("synapter")


After installation, r Biocpkg("synapter") will have to be explicitly loaded with

library("synapter")


so that all the package's functionality is available to the user.

# Data analysis using r Biocpkg("synapter"){#sec:analysis}

## Preparing the input data{#sec:input}

Preparation of the data for r Biocpkg("synapter") requires the .raw data first to be processed with Waters' ProteinLynx Global Serve (PLGS) software. The PLGS result is then exported as csv spreadsheet files in user specified folders. These csv files can then be used as input for r Biocpkg("synapter").

We also highly recommend users to acquaint themselves with the PLGS search algorithm for data independent acquisitions [@PLGS2009].

First the user has to specify the output folders for files to be used in synapter analysis as demonstrated in the figures. After the folders are specified ignore the message that appears requiring restarting PLGS.

{#fig:plgs1} {#fig:plgs2} {#fig:plgs3}

At the first stage PLGS performs noise reduction and centroiding, based on user specified preferences called processing parameters. These preferences determine thresholds in intensity for discriminating between noise peaks and peptide and fragment ion peaks in high and low energy functions of an acquisition. The optimal value of these parameters is sample dependant and different between MS$^E$ and HDMS$^E$ modes. For r Biocpkg("synapter") to function properly all acquisitions in the analysis have to be processed with the same thresholds, optimal for the mode identifications are transferred from (typically HDMS$^E$ mode). The user is expected to identify optimal parameters himself for every new sample type by repeatedly analysing a representative acquisition with different thresholds.

After the ions peaks have been determined and centroided, the ions representing charge states and isotopes of a peptide are collapsed into a single entity called EMRT (exact mass retention time pair). The EMRTs in low energy function represent unidentified peptides and are assigned peptides sequences during database search. The total list of EMRTs can be found in the pep3DAMRT.csv file and it is one of the r Biocpkg("synapter") input files for the runs used for quantitation (typically MS$^E$ mode)

Prior to the database search, randomised entries are added to the database to allow PLGS to compute protein false positive rate. The randomised entries can either be added automatically or manually, using the Randomise Databank function in the Databank admin tool. To properly prepare the files for r Biocpkg("synapter"), the user has to add randomised entries manually via Databank admin tool, since only then randomised entries identified in the database search will be displayed in the csv output. The following figures demonstrate how to create a randomised databank manually using one randomised entry per regular entry.

{#fig:plgs4} {#fig:plgs5}

The user is also expected to use a minimum of 1 fragment per peptide, 3 fragments per protein and 1 peptide per protein identification thresholds and 100\% False Positive Rate^[This is erroneously termed false positive rate in the software and manuscript and should be considered a false discovery rate.] for protein identification during database search for all of the acquisitions in the analysis as demonstrated in figure. This allows to maximise the number of identified peptides from the randomised part of the database, needed to estimate peptide identifications statistics. The total list of identified peptides is given in final_peptide.csv files. A single final_peptide.csv file has to be supplied to r Biocpkg("synapter") for every run in the analysis (for both identification and quantitation runs).

More details and screenshots are available in a separate document available at https://lgatto.github.com/synapter/.

## HDMS$^E$/MS$^E$ data analysis{#sec:synapter}

The analysis of pairs of HDMS$^E$ and MS$^E$ data files is based on the following rationale -- combine strengths or each approach by matching high quality HDMS$^E$ identifications to quantified MS$^E$ EMRTs applying the following algorithm:

1. Apply various peptide filters to HDMS$^E$ and MS$^E$ peptides to obtain two sets of reliably identified unique proteotypic peptides.
2. Use shared HDMS$^E$ and MS$^E$ peptides to model the deviations in retention time between the two mass spectrometer runs.
3. Optimise the parameters that will be used to optimally match all HDMS$^E$ peptides and quantified MS$^E$ EMRTs using a grid search.
4. Using the best parameters, match identified HDMS$^E$ peptides to quantified MS$^E$ EMRTs.

## Different pipelines{#sec:pipelines}

Two different pipeline are available to the user:

### Wrapper function

The synergise function is a high level wrapper that implements a suggested analysis to combine two files (see next paragraph for details). A set of parameters can be passed, although sensible defaults are provided. While the analysis is executed, a html report is created, including all result files in text spreadsheet (csv format) and binary R output. This level allows easy scripting for automated batch analysis. Using data from the r Biocexptpkg("synapterdata") package, the following code chunk illustrates the synergise usage. An example report can be found online at https://lgatto.github.com/synapter/.

library("synapterdata")
hdmsefile <- getHDMSeFinalPeptide()[2]
basename(hdmsefile)
msefile <- getMSeFinalPeptide()[2]
basename(msefile)
msepep3dfile <- getMSePep3D()[2]
basename(msepep3dfile)
fas <- getFasta()
basename(fas)
## the synergise input is a (named) list of filenames
input <- list(identpeptide = hdmsefile,
quantpeptide = msefile,
quantpep3d = msepep3dfile,
fasta = fas)
## a report and result files will be stored
## in the 'output' directory
output <- tempdir()
output

res <- synergise(filenames = input, outputdir = output)

performance(res)


See ?synergise for details.

### Detailed step-by-step analysis

The user can have detailed control on each step of the analysis by executing each low-level function manually. This pipeline, including generation of data containers (class instances) and all available operations are available in ?Synapter. This strategy allows the maximum flexibility to develop new unexplored approaches.

## Using master peptide files{#sec:master}

While analysing one MS$^E$ file against one single HDMS$^E$ file increased the total number of reliably identified and quantified features compared to each single MS$^E$ analysis, a better procedure can be applied when replicates are available. Consider the following design with two pairs of files: HDMS$^E_1$, MS$^E_1$, HDMS$^E_2$ and MS$^E_2$. The classical approach would lead to combining for example, HDMS$^E_1$ and MS$^E_1$ and HDMS$^E_2$ and MS$^E_2$. However, HDMS$^E_1$ -- MS$^E_2$ and HDMS$^E_2$ -- MS$^E_1$ would also be suitable, possibly leading to new identified and quantified features. Instead of repeating all possible combinations, which could hardly be applied for more replicates, we allow to merge HDMS$^E_1$ and HDMS$^E_2$ into a new master HDMS$^E_{12}$ and then using it to transfer identification to both MS$^E$ runs. In addition to leading to a simpler set of analyses, this approach also allows to control the false positive rate during the HDMS$^E$ merging (see section Choosing HDMS$^E$ files). Such master HDMS$^E$ files can be readily created with the makeMaster function, as described in section Generating a master file.

We will use data from the r Biocexptpkg("synapterdata") to illustrate how to create master files.

### Choosing which HDMS$^E$ files to combine{#sec:whichhdmse}

In a more complex design, a greater number of HDMS$^E$ files might need to be combined. When combining files, one also accumulates false peptides assignments. The extent to which combining files increases new reliable identification at the cost of accumulating false assignments can be estimated with the estimateMasterFdr function.

To illustrate how FDR is estimated for master HDMS$^E$ files, let's consider two extreme cases.

• In the first one, the two files (each with $1000$ peptides filtered at an FDR of $0.01$) to be combined are nearly identical, sharing $900$ peptides. The combined data will have $900 (shared) + 2 \times 100 (unique)$ peptides and each file, taken separately is estimated to have $1000 \times 0.01 = 10$ false positive identifications. We thus estimate the upper FDR bound after merging the two files to be $\frac{20}{1100} = 0.0182$.

• In the second hypothetical case, the two files (again each with $1000$ peptides filtered at a FDR of $0.01$) to be combined are very different and share only $100$ peptides. The combined data will have $100 (shared) + 2 \times 900 (unique)$ peptides and, as above, each file is estimated to have $10$ false discoveries. In this case, we obtain an upper FDR bound of $\frac{20}{1900} = 0.0105$.

In general, the final false discovery for two files will be $$FDR_{master} = \frac{nfd_{1} + nfd_{2}}{union(peptides~HDMS^E_{1}, peptides~HDMSE_{2})}$$

where $nfd_{i}$ is the number of false discoveries in HDMS$^E$ file $i$. Note that we do not make any assumptions about repeated identification in multiple files here.

estimateMasterFdr generalised this for any number of HDMS$^E$ files and indicates the best combination at a fixed user-specified masterFdr level. Mandatory input is a list of HDMS$^E$ file names and a fasta database file name to filter non-unique proteotypic peptides.

inputfiles <- getHDMSeFinalPeptide()


The result of estimateMasterFdr stores the number of unique proteotypic peptides and FDR for all possible r sum(choose(length(inputfiles),2:length(inputfiles))) combinations of r length(inputfiles) files. A summary can be printed on the console or plotted with plot(cmb) (see figure 1).

## using the full set of 6 HDMSe files and a
## fasta database from the synapterdata package
inputfiles <- getHDMSeFinalPeptide()
fasta <- getFasta()
cmb <- estimateMasterFdr(inputfiles, fasta, masterFdr = 0.02)
cmb

plot(cmb)


The best combination can be extracted with the bestComb function.

bestComb(cmb)


See ?estimateMasterFdr and references therein for more details about the function and the returned object.

### Generating a master file{#sec:makeMaster}

Now that we have identified which files should be used to create the master file, we can directly pass the relevant identification files to the makeMaster function to generate the master file. The function has one mandatory input parameter, pepfiles, a list oh identification file names to be merged. The output is an object of class MasterPeptides that stores the relevant peptides from the original input files. The result can be saved to disk using saveRDS for further analysis, as described in section HDMS$^E$/MS$^E$ data analysis.

master <- makeMaster(inputfiles[bestComb(cmb)])
master


More details can be found in the function documentation accessible with ?makeMaster.

## Summary

Two functions are needed to choose a set of identification run files and create the master identification run. One function enables to perform a complete identification transfer. The following table summarises all there is to know to utilise r Biocpkg("synapter")'s functionality.

Function Description

synergise Runs the complete identification transfer estimateMasterFdr Chooses which files to be used to create the master IR makeMaster Creates the master IR

# Analysing complete experiments{#sec:msnbase}

The functionality described in this section relies on the r Biocpkg("MSnbase") package [@Gatto2012], which is installed by default with r Biocpkg("synapter"). Please refer to the r Biocpkg("MSnbase") Bioconductor web page. the associated vignettes and the respective manual pages for more details.

The r Biocexptpkg("synapterdata") already provides preprocessed PLGS data. Six Synapter instances are available: 3 replicates (labelled a, b and c) of the Universal Proteomics Standard (UPS1) 48 protein mix at 25 fmol and 3 replicates at 50 fmol, in a constant E. coli background. The 6 files can be loaded in your working space with

data(ups25a, ups25b, ups25c, ups50a, ups50b, ups50c)


## Applying the Top 3 approach{#sec:top3}

We will start by describing the analysis of ups25a in details, and then show how to analyse all the runs using more compact code. The first step of our analysis is to convert the r Biocpkg("synapter") object output (an Synapter instance), into a r Biocpkg("MSnbase")-compatible object, called an MSnSet, that we will name ms25a. We can obtain a description of the MSnSet object by typing its name.

ms25a <- as(ups25a, "MSnSet")
class(ups25a)[1]
class(ms25a)[1]
ms25a


It contains quantitation information about r nrow(ms25a) peptides for r ncol(ms25a) sample. In the code chunk below, we update the default sample name Synapter1 with a more meaningful one.

sampleNames(ms25a)
sampleNames(ms25a) <- "ups25a"
sampleNames(ms25a)


Quantitative data and meta-data, which has been acquired by r Biocpkg("synapter"), can be extracted with the exprs and fData methods.

tail(exprs(ms25a))
tail(fData(ms25a)[, c(2,9)])
fvarLabels(ms25a)


We will describe a how to process the data using a Top 3 approach, where the 3 most intense peptides of each protein are used to compute the protein intensity, using the topN and combineFeatures methods. The former allows to extract the top most intense peptides (default n is 3) and remove all other peptides from the MSnSet object. The latter than aggregates the n most intense peptides per protein using a user-defined function (sum, below). Finally, we also scale protein intensity values depending on the actual number of peptides that have summed. This number of quantified peptides can be calculated (after topN, but before combineFeatures) with nQuants.

ms25a <- topN(ms25a,
groupBy = fData(ms25a)$protein.Accession, n = 3) nPeps <- nQuants(ms25a, groupBy = fData(ms25a)$protein.Accession)
ms25a <- combineFeatures(ms25a,
fData(ms25a)$protein.Accession, "sum", na.rm = TRUE, verbose = FALSE) head(exprs(ms25a)) head(nPeps) ## scale intensities exprs(ms25a) <- exprs(ms25a) * (3/nPeps) ## NaN result from the division by 0, when ## no peptide was found for that protein head(exprs(ms25a))  ## Batch processing{#sec:batchprocessing} The code chunk below repeats the above processing for the other 5 UPS1/E. coli runs. nms <- c(paste0("ups", 25, c("b", "c")), paste0("ups", 50, c("a", "b", "c"))) tmp <- sapply(nms, function(.ups) { cat("Processing", .ups, "... ") ## get the object from workspace and convert to MSnSet x <- get(.ups, envir = .GlobalEnv) x <- as(x, "MSnSet") sampleNames(x) <- .ups ## extract top 3 peptides x <- topN(x, groupBy = fData(x)$protein.Accession, n = 3)
## calculate the number of peptides that are available
nPeps <- nQuants(x, fData(x)$protein.Accession) ## sum top3 peptides into protein quantitation x <- combineFeatures(x, fData(x)$protein.Accession,
"sum", na.rm = TRUE, verbose = FALSE)
## adjust protein intensity based on actual number of top peptides
exprs(x) <- exprs(x) * (3/nPeps)
## adjust feature variable names for combine
x <- updateFvarLabels(x, .ups)
## save the new MSnExp instance in the workspace
varnm <- sub("ups", "ms", .ups)
assign(varnm, x, envir = .GlobalEnv)
cat("done\n")
})


We now have 6 MSnSet instances, containing protein quantitation for the 6 UPS/E. coli runs.

## Combining data and filtering{#combine}

We now want to filter data out based on missing quantitation data, retaining proteins that have been quantified in at a least two out of three replicates. Filtering based on missing data can be done using the filterNA method and a maximum missing data content as defined by pNA. Multiple MSnSet instances can be combined with the combine method, which is described in details in the MSnbase-demo vignette^[The vignette is accessible from within R with vignette("MSnbase-demo" package = "MSnbase").]. The 6 objects have appropriate distinct sample names and common feature (protein) names, which will be used to properly combine the quantitation data.

ms25 <- combine(ms25a, ms25b)
ms25 <- combine(ms25, ms25c)
dim(ms25)
ms25 <- filterNA(ms25, pNA = 1/3)
dim(ms25)


Once combined and filtered, the 25 fmol group retains r nrow(ms25) entries with at least 2 out of 3 quantitation values, out of the r nrow(ms25a) total number of proteins.

ms50 <- combine(ms50a, ms50b)
ms50 <- combine(ms50, ms50c)
dim(ms50)
ms50 <- filterNA(ms50, pNA = 1/3)
dim(ms50)


Similarly, the 50 fmol group retains r nrow(ms50) entries with at least 2 out of 3 quantitation values, out of the r nrow(ms50a) initial proteins.

We now combine the two subgroups into one MSnSet object that contains all 6 samples and filter for proteins that are observed in both groups, i.e retaining proteins with a maximum of 2/6 missing values. We also compute a summary table with the number of protein that have 4, 5, or 6 quantitation values across the 6 samples.

msUps <- combine(ms25, ms50)
msUps <- filterNA(msUps, pNA = 2/6)
table(apply(exprs(msUps), 1,
function(.x)sum(!is.na(.x))))


We obtain a final data set containing r nrow(msUps) proteins. Finally, we normalise protein intensities in each sample to correct for experimental loading biases and pipetting errors. To do so, we compute 6 sample medians using all constant E. coli background proteins and divide each protein by its respective sample mean.

ecoli <- -grep("ups$", featureNames(msUps)) meds <- apply(exprs(msUps)[ecoli, ], 2, median, na.rm=TRUE) exprs(msUps) <- t(apply(exprs(msUps), 1, "/", meds))  This same procedure could be applied with a set of constant spikes to estimate absolute protein quantities. ## Statistical analysis of differentially expressed proteins The UPS1 spiked-in protein mix is composed of 48 proteins, r length(grep("ups", featureNames(msUps))) of which have been observed and quantified in our final data object. In this section, we will illustrate how to analyse the r nrow(msUps) proteins to extract those that show differences between the two groups and show that these candidates represent the UPS1 spikes. The R environment and many of the available packages allow extremely powerful statistical analysis. In this document, we will apply a standard t-test on$log_2$transformed data for convenience, to calculate p-value for individual proteins (pv variable below). For best performance with small number of samples and more complex designs, we recommend the Bioconductor r Biocpkg("limma") package [@limma]. We then perform multiple comparison adjustment using the qvalue from the r Biocpkg("qvalue") package, that implements the method from [@Storey2003] (qv variable below). The r Biocpkg("multtest") package provides several other p-value adjustment methods. We will also compute$log_2$fold-changes and illustrate the results on a volcano plot (figure 2). Figure 3 illustrates the UPS1 proteins and samples on a classical heatmap. ## use log2 data for t-test exprs(msUps) <- log2(exprs(msUps)) ## apply a t-test and extract the p-value pv <- apply(exprs(msUps), 1 , function(x)t.test(x[1:3], x[4:6])$p.value)
## calculate q-values
library("qvalue")
qv <- qvalue(pv)$qvalues ## calculate log2 fold-changes lfc <- apply(exprs(msUps), 1 , function(x) mean(x[1:3], na.rm=TRUE)-mean(x[4:6], na.rm=TRUE)) ## create a summary table res <- data.frame(cbind(exprs(msUps), pv, qv, lfc)) ## reorder based on q-values res <- res[order(res$qv), ]


In the above example, quantitation values and statistics data are saved in a summary dataframe (res), that can be exported to a comma-separated spreadsheet with

write.csv(res, file = "upsResults.csv")


r$of its q-value along the y axis. Proteins with large fold-changes are positioned on the sides of the plot, while proteins with low q-values are at the top of the figure. The most promising candidates are this located on the top corners. In our case, the UPS proteins (in blue) have$log2$fold-changes around -1 (vertical dotted line), as expected. The horizontal dotted line represents a q-value threshold of 0.10."} plot(res$lfc, -log10(res$qv), col = ifelse(grepl("ups$", rownames(res)), "#4582B3AA", "#A1A1A180"), pch = 19, xlab = expression(log[2]~fold-change), ylab = expression(-log[10]~q-value)) grid() abline(v = -1, lty = "dotted") abline(h = -log10(0.1), lty = "dotted") legend("topright", c("UPS", "E. coli"), col = c("#4582B3AA", "#A1A1A1AA"), pch = 19, bty = "n")

r
heatmap(exprs(msUps)[grep("ups",featureNames(msUps)), ])

sel1 <- (res\$qv < 0.1)
signms <- rownames(res)[sel1]
i <- grep("ups", signms)
n1 <- length(i)
n2 <- length(signms) - n1

k <- grep("ups", rownames(res))
knitr::kable(round(res[k, ], 3),
caption = "UPS1 proteins.")


A total r n1 proteins show a statistically different pattern between the two groups, at a false discovery rate of 10\%. Table upstab summarises the results for all UPS1 proteins.

# Session information{#sec:sessionInfo}

All software and respective versions used to produce this document are listed below.

sessionInfo()


# References

## Try the synapter package in your browser

Any scripts or data that you put into this service are public.

synapter documentation built on May 24, 2018, 6:01 p.m.