==== Differential protein expression (with mass spec) ====

This method is for label-free samples from our Proteomics Core Facility, which has some [[http://massspec.wi.mit.edu/documents/Scaffoldhowto060513.pdf|Scaffold quick instructions]].

  * Using the (free) Scaffold Viewer (available from [[http://www.proteomesoftware.com/products/free-viewer|Proteome Software]], with a large [[http://www.proteomesoftware.com/pdf/scaffold_users_guide.pdf|User's Manual]]), open the sf3 file.

  * By default, three filters prevent all mapped proteins from being displayed:
    * Protein Threshold (default = 99%)
    * Min # Peptides (default = 2)
    * Peptide Threshold (default = 95%) 
  * These filters are typical good with the default settings.

  * Scaffold has multiple display and quantification options.
    * From the Scaffold User's Manual:
      * Spectrum Counting methods are the most reliable in answering the question, "Is anything changing between experimental conditions?".
      * Precursor Ion Intensity quantification methods are very reliable in answering the question, "How much is the amount of change I am dealing with?"
      * The Total Ion Count (TIC) methods can answer both questions but not very well. 
    * For a quick QC check: 
      * Look at the first few most highly expressed proteins.  
      * Are they within 2-fold or so of each other?  
      * If some samples have much lower or higher counts, their sensitivity may differ so much that samples may not be comparable.

    * Under Display Options, select "Total Spectrum Count"

    * For Spectrum Counting, click on the "Quantitative Analysis" icon (showing a bar graph).
       * Keep "Use Normalization" checked.
       * For Quantitative Method, select "Total Spectra".
       * Click the Apply button.
       * Next to Display Option: select Quantitative Value (Normalized Total Spectra)
       * Export (top menu) => Current View.
       * Use this file to identify differentially expressed proteins (to be explained below).

    * For Precursor Ion Intensity quantification, click on the "Quantitative Analysis" icon (showing a bar graph).
       * Keep "Use Normalization" checked.
       * For Quantitative Method, select "Top 3 Precursor Intensity".
       * Click the Apply button.
       * Next to Display Option: select Quantitative Value (Normalized Top 3 Precursor Intensity)
       * Export (top menu) => Current View.
       * Use this file for visualization (to be explained below).
  
    * To identify differentially expressed proteins, use the normalized Total Spectrum Count
      * Recommended statistic: t-test on log2 transformed values
      * Correct p-values with FDR (or an alternate method) 

    * For pathway analysis: [[https://david.ncifcrf.gov/|DAVID]] usually works fine.

    * For visualization:
      * Draw a heatmap (Cluster3.0 -> Java TreeView) using the normalized Top 3 Precursor Intensities. 
      * Draw scatterplot using the normalized Top 3 Precursor Intensities, highlighting the differentially expressed proteins (from the Total Spectrum).  


==== Recommendations from Northeastern (May Institute, Vitek Lab)  ====

  * Best input is peptide-level "peak intensities", which are any continuous metric, such as Scaffold's
    * Average Precursor Intensity
    * Total Precursor Intensity
    * Top Three Precursor Intensities
  * Ideal analysis pipeline is to input these values into MSstats for pre-processing, statistics, and data visualization
  * Preprocessing steps recommended by (performed by) MSstats:
    * Log2 transform
    * Median-normalize across samples and runs (ignoring any 0s)
    * Convert all 0s to NA
    * Censor low measurements
      * Get median
      * Get 99.9th (or other percentile) to identify right tail of distribution ("r")
      * Get threshold of left side ("l") of the distribution (2*median - r)
      * Censor all values less than "l"
    * Impute all missing values using a MNAR method, such as the accelerated failure model
    * Summarize all features of a protein using Tukey's median polish (TMP), but ignore proteins with only 1 peptide (or risk increased false positive rate)
  * Model each protein with a linear mixed-effects model
     * Limma does a good job too, but it doesn't handle all the experimental designed handled by MSstats
   * Use model to calculate fold changes and raw p-values
   * Correct all p-values with FDR (BH method)
   * Draw summary plots (volcano plots, MA plots) 

==== Preparing and processing an experiment with MSstats  ====

  * Export peptide-level intensity values from your favorite MS quantification software.

  * Create a peptide intensity file
      * Organize the dataset so the first three columns are Gene.symbol, Protein.Accession, and Peptide.sequence
        * If needed, convert each protein accession to a gene symbol
        * Replace any intensities shown as "-" with 0
        * If Excel is used, check that gene symbols aren't being converted to dates
      * After the first 3 columns, the remaining columns hold intensities, one column per sample 

  * To avoid losing information about peptides that could have originated from multiple proteins and/or genes, merge peptide rows representing more than 1 protein and/or gene
     * Sample command: sort -k3,3 Peptide_intensities.matrix.txt | groupBy -g 3 -c 1,2,3,4,5,6,7,8,9,10 -o distinct >| Peptide_intensities.matrix.mergedByPeptide.txt
     * Make sure that all rows of the output file are unique.

  * Create a sample description file
     * Columns are  Run, Condition, BioReplicate
     * See MSstats documentation on how to use these fields to represent technical replicates, biological replicates, and paired designs.
     * Replication is not required for subsequent protein quantification but is required for statistical analysis.
 
  * Run MSstats using peptide intensities and sample description as input files.
    * For sample code, see **/nfs/BaRC_code/R/analyze_MS_with_MSstats/analyze_MS_with_MSstats.R**

==== References  ====

  * Choi et al., 2017 [[https://pubs.acs.org/doi/abs/10.1021/acs.jproteome.6b00881|ABRF Proteome Informatics Research Group (iPRG) 2015 Study: Detection of Differentially Abundant Proteins in Label-Free Quantitative LC−MS/MS Experiments]]
  * MSstats at [[https://bioconductor.org/packages/release/bioc/html/MSstats.html|Bioconductor]] and [[https://github.com/MeenaChoi/MSstats|GitHub]]
  * Other references on these topics
    * Lazar et. al., 2016 [[https://pubs.acs.org/doi/full/10.1021/acs.jproteome.5b00981|Accounting for the Multiple Natures of Missing Values in Label-Free Quantitative Proteomics Data Sets to Compare Imputation Strategies]]
    * Wei et al., 2018  [[https://www.nature.com/articles/s41598-017-19120-0|Missing Value Imputation Approach for Mass Spectrometry-based Metabolomics Data]]