Thesis-MS-data-processing

Overview of the MS data processing process. GUIDE - READ BEFORE PROCESSING Content This guide goes through the set up for processing from individual .mzML files to a a combined TIC sheet It will go through the following aspects of processing:

1. Workspace What does the workspace need to contain?
2. Converting ‘.RAW’ to ‘.mzML’ What will be assumed about your data
3. Metadata The format it will need to be in 4.Processing settings
4. Sense checks

---

1. Workspace Do not work from the shared drive R-processing folder, create a new directory in your workspace. i.e. G:/Shared drives/Mass Spec _Metabolomics/James/NewExperimentProcessing In this folder you will need: A copy of the 'DI-ESI Processing Script' R-File Your .mzML files in a folder called 'RAW' An empty folder called 'SUM' A 'Metadata.csv' file with your sample info on

Packages & R version This script was written in R 4.10, newer versions may have changes that aren't compatible, so if there is a recurring error consider reverting back to this version. Packages will take some time to install, but you should only have to do it once (they will be saved to your library). During the installation process you may be asked if you want to install updates in the console, this may alter the functions in later versions so sticking with ‘n’ is a safe bet. Once installed you can just library them rather than having to reinstall To this end the script has the installation code greyed out. You can remove the ‘#’ from the front and the code can be run if needed. Otherwise run the > lapply(c(“MSnbase”..... Code and it will library the required packages, and print ‘TRUE’ in the console if successful If FALSE appears, reinstall the corresponding package Firstly will need some mass spec data processing specific packages from a repository known as BiocManager (version 3.10 - later versions may work but this is what it was written with) MSnbase MALDIquant MALDIquantForeign Dplyr (If you have a problem with the TechReps portion of the code it is likely because you haven't got dplyr installed and in the library)

---

2. Converting ‘.RAW’ to ‘.mzML’ Prior to processing you need to convert your '.raw' files output by the mass spec into '.mzML' files This can be done using MSConvert (Proteowizard) Here you can also cut out prewash/wash scans from your files i.e. if your sample peak in the chromatogram is between scans 20 and 40, you can select only these scans Use the subset filter to do this, don't forget to press 'ADD' Check your .mzML files after they've been converted using SeeMS, look for: Should only have your selected scans in (although the entire chromatogram image will remain) Should see the mass spectra seen in mass lynx on the actual MS computer Should all be similar size files (if there are some particular small/big ones double check them as it occasionally errors) What will be assumed about your data The script will assume that every .mzML file in the raw folder wants processing, and that it is all in the metadata. If it is not in the metadata it will stop running after the scan-summing step as the list of files will not match up. It will assume that all the scans in the .mzML files are wanted and will contribute to the end value If you only want a select scan range, specify this in the conversion step.

---

3. Metadata You will need a Metadata.csv file with all your sample information in it. It will need the first three columns to be: File [This is the file name the data is saved under] SampleID [This is the samples unique identifier] Techrep [If only one taken still call it ‘1’] After this you can add any extra information you want and give it whatever header you want [If you do add new columns don't leave any file rows with blank cells] Note: If you have acquired technical reps, the output will give you a table of masses in each rep and a separate table where technical reps are averaged.

---

4. Processing Settings There are two sections of the processing that we can edit to fit our needs. Below will go through the basic considerations. NOTE: This script is written to give a very high resolution detection rate (good for noisy samples with low intensity) and a standard [low] resolution detection rate (Good for most normal experiments). Processing steps basic guide: A. Transformation B. Smoothing C. Baseline correction D. Normalisation E. Peak alignment F. Peak detection G. Peak binning

A. Transformation Reasoning for transformation: To stabilise the variance in the data The detected intensity of an analyte can vary due to many factors in the detection process, meaning that there can be considerable variation between peaks. By normalising the peaks we can make them more comparable, and give more accurate and precise measurements Types of transformations available and which to use: Log transformation {“log”, “log2”,“log10”} Good for removing the impact of huge outliers Often over corrects, leaving too much noise Square root transformation {“sqrt”} Better choice for normally distributed data This is the default in the script

B. Smoothing Reasoning for smoothing: To reduce the noise in the data This makes it easier to identify actual peaks HOWEVER overcorrection can smooth out small peaks or closely separated peaks. For example, if you're looking for two different metabolites at 191.0152 and 191.0538 these masses can be merged into one peak, giving one combined value Types of smoothing available and which to use: SavitzkyGolay {"SavitzkyGolay"} Uses polynomial function to smooth the peaks Higher accuracy, because not only neighbouring peaks taken into account Good for preserving peak shape MovingAverage {"MovingAverage"} Uses neighbouring peaks to smooth the peak Lover accuracy but less computationally expensive Does not preserve the shapes of the peaks Other parameters Half window size (HWS) Determines how many data points are included in the smoothing Larger value = more smoothing If you’re trying to remove noise from your data a larger HWS Default in script is set to 4

C. Baseline correction Reasoning for baseline correction: It removes the background noise from the spectra. Noise from the instrument/matrix/sample/etc will always be present Removing it makes the real peaks more obvious Artefacts due to noise are removed so you don't get false peaks The quantification/calibration of the peaks intensity is more accurate as it is being measured from the same baseline Peaks starting on a higher baseline may incorrectly be considered higher intensity, by making sure the baseline is corrected this doesn't happen. Methods of baseline correction available and which to use: Statistics-sensitive Non-linear Iterative Peak-clipping algorithm {"SNIP"} Uses basic threshold to remove baseline noise Threshold slightly above the noise ratio of the data Little chance of introducing artefacts into the data TopHat {"TopHat"} Uses moving average to remove baseline noise The neighbouring data has most influence on whether a point is removed, meaning it can falsely add artefacts if there is a constant signal More effective at removing noise but can introduce artefacts ConvexHull {"ConvexHull"} Removes a convex hull, more effective than above options Good at flattening consistent baseline shifts But not good at removing baselines that aren't smooth Median {"median"} Removes the median of the data More effective than convex hull on data that isn't smooth But still may introduce artefacts Other parameters Iterations (when using SNIP) / halfWindowSize (when using TopHat) The higher the iterations/HWS the more points included in the smoothing process For high resolution data this value may need to be lowered Default for the script is 250

D. Normalisation/calibration Reasoning for normalisation: Signal intensity can be fairly arbitrary, so we need to normalise it into a more usable format TIC is easier to work with and more comparable between samples where average intensity is not. Types of normalisation available and which to use: Total Ion Current {"TIC"} This normalises the intensity data to the entire spectrum, so that each peak’s intensity is a percentage of the overall signal Probabilistic Quotient Normalization {"PQN"} This divides the value for each peak by the signal of its precursor ion (the ion responsible for the peak) The problem is it is majorly computationally expensive It is more accurate though as it will prevent over/under representation of certain compounds. Median normalisation {"median"} Uses the median of the data to normalise to. Less accurate but better at dealing with outliers.

E. Peak alignment Reasoning for peak alignment: The detection of peaks to such fine masses means that there will be discrepancies between the m/z values for different samples. These discrepancies between samples would mean inaccurate calling of a peak's intensity, when they are the same peak Essentially it makes sure you are comparing the same peaks. Parameters which to use: Warping method = “Lowess” Works well on noisy data and data with outliers Based on local density of points There is the option for quadratic, linear and cubic but Lowess works best HalfWindowSize = 5 Similar to before it determines how many points are included in the ‘peak’ If you want higher resolution (aka more separation) you can decrease this number Signal to Noise ratio (SNR) = 3 This separates peaks from noise A local maximum (data point) has to be higher than this ratio to be considered a peak Lowering this value will lower the threshold for what is called a peak, meaning low intensity data is picked up Tolerance = 1e-4 This is the local threshold for what is an identical peak. Default is set to 0.0001 (4 decimal places)

F. Peak detection Reasoning for peak detection: This is where actual peaks are selected instead of signal curves We want the intensity value of a single m/z value not the entire peak shape. Here we can significantly impact the resolution of our output by changing how the peaks are detected. This is where the high resolution or low resolution settings come in. Types of peak detection available and which to use: Mean absolute deviation {“MAD”} It uses the median data points to create a threshold, and then anything above this threshold is considered a peak and not noise. Using the median of the data makes this test more robust as outliers don’t affect the outcome.
Friedman's Super Smoother {“SuperSmoother”} Uses a moving average to work out what a peak is and what is the base line Less robust as outliers can impact the average Parameters to consider: HalfWindowSize = [Low Resolution] 10, [High resolution] 3 Same as before this is how many neighbouring points are included in the assessment of a peak. Decreasing the value makes the peak detection more specific. Signal to noise ratio (SNR) = [Low Resolution] 4, [High resolution] 2 Lower value will call more peaks, higher the opposite refineMZ = “descendPeak” This defines how a peak is identified descendPeak sees the area of the peak as from the tip down until the measured signal increases again (aka the triangle of the peak) Closely positioned peaks that have been over-smoothed will be merged into one data point as there will be no upward angle of the second peak This can be an issue worth checking Signal percentage = 50 This determines how much of the peak area contributed to the m/z value calculated The default has it set to 50%, so half the peak area determines the value

G. Peak binning Reasoning for peak binning: This is for aligning all the m/z values so that there isn't discrepancies across the entire sample sets It is unifying the mz value across all samples Types of peak binning available and which to use: Strict {“strict”} Creates bins that cannot contain 2 or more peaks in the same sample Provides more distinct peaks Relaxed {“relaxed”} Allows multiple peaks in the same bin Can capture broad peaks better Binning tolerance = [Low Resolution] 0.05, [High resolution] 0.005 Determines how big the bin will be Smaller value = more bins (and with strict, more possible peaks)

---

5. Sense check At the bottom of the script there is a list of plots that can be printed to check if the processing is doing what you expect it to. You will need to input a mass range of an expected metabolite in your sample. I chose 190.9-191.1 because I was expecting 2 peaks in my data in this range, and both should be present in all samples. Print the plots one by one and see how they change. You can edit your parameters at the top of the script to see how it changes things, but remember to reprocess the sections you are testing in the main body of the processing (you can just click through it all again) There is also a parameters output which will give you the m/z levels detected using the settings chosen.