📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-workflows-proteomics-pipeline / references / usage-guide.md

Proteomics Pipeline Usage Guide

Overview

End-to-end workflow for label-free proteomics analysis from MaxQuant/DIA-NN output to differential protein abundance.

Prerequisites

BiocManager::install(c('limma', 'DEqMS', 'proDA', 'MSstats', 'MSstatsTMT', 'ashr'))
install.packages(c('pheatmap', 'ggplot2', 'arrow', 'iq'))

Quick Start

Tell your AI agent what you want to do: - "Run the proteomics pipeline on my MaxQuant output" - "Find differentially expressed proteins between conditions" - "Process my DIA-NN results and run differential analysis"

Example Prompts

Basic Analysis

"I have proteinGroups.txt from MaxQuant, run the full pipeline"

"Normalize my proteomics data and find differential proteins"

QC and Preprocessing

"Check sample quality with PCA and correlation heatmap"

"Handle missing values correctly by modeling the dropout rather than imputing"

Differential Analysis

"Run limma to find proteins changed between treatment and control"

"Use MSstats for differential analysis with my peptide-level data"

Pipeline Stages

1. Data Import

Load proteinGroups.txt (MaxQuant) or report.parquet (DIA-NN 1.9+)
Filter contaminants, reverse, and only-identified-by-site BEFORE normalizing
Extract intensity columns (LFQ intensity, not raw Intensity, for between-sample work)

2. Transformation

Replace 0 with NA
Log2 transform
Median centering normalization

3. Completeness Filtering

Keep proteins valid in >= ~60% of replicates in at least one condition
Filter before any missing-value handling, so only shallow gaps remain

4. Missing-Value Handling

Preferred: MODEL the left-censored MNAR dropout (proDA / msqrob2 / MSstats-AFT); no imputation
Fallback only: downshift imputation, which manufactures false positives for on/off proteins near the detection limit
Report a protein fully missing in one group as "undetected", not as a fold change

5. Quality Control

PCA: Check replicate clustering
Correlation heatmap: Sample similarity
Missing value patterns: Random or systematic

6. Differential Analysis

limma: empirical-Bayes moderated t-test (trend=TRUE, robust=TRUE); treat() for a minimum fold change
DEqMS: count-aware moderation when quantification depth varies
proDA: probabilistic dropout model (no imputation)
MSstats / MSstatsTMT: feature-level mixed models, plus the IRS bridge for multi-batch TMT

7. Output

Differential proteins table
Volcano plot
Heatmap of significant proteins

Input Requirements

MaxQuant Output

proteinGroups.txt  # Protein-level quantification
evidence.txt       # Peptide-level (for MSstats)
annotation.csv     # Sample metadata

Sample Annotation

sample,condition,replicate
Sample1,Control,1
Sample2,Control,2
Sample3,Treatment,1
Sample4,Treatment,2

Expected Outputs

File	Description
proteomics_results.csv	All proteins with statistics
proteomics_results_volcano.pdf	Log2FC vs -log10(p-value)
proteomics_results_raw_boxplot.pdf	Raw per-sample log2 distributions, inspected before normalization
proteomics_results_pca.pdf	Sample clustering
proteomics_results_heatmap.pdf	Significant proteins

Typical Results

2000-5000 quantified proteins (cell lysate)
50-500 differential proteins (10%)
Fold changes typically 1.5-4x

Tips

Missing values: model the MNAR dropout (proDA/msqrob2/MSstats-AFT) rather than impute; downshift imputation manufactures systematic false positives
Normalization: median centering is the default for balanced designs; never normalize an AP-MS/enrichment pulldown this way
Completeness: filter on per-group completeness before any missing-value handling, not a blanket >50% rule
Replicates: minimum 3 biological replicates per condition
Contaminants: filter MaxQuant contaminants, reverse, and only-identified-by-site BEFORE log + normalize