Proteomics QC - Usage Guide

Overview

Quality control for bottom-up proteomics framed as a three-level funnel: instrument/raw-signal QC (mass accuracy, RT/iRT fit, FWHM, TIC vs injection time, % MS2 identified), identification/run QC (missed cleavages, charge states, sample-handling PTM artifacts, contaminants), and experiment/quantitative QC (replicate correlation on log2, CV on the linear scale, completeness, missingness mechanism, PCA/batch, TMT channel balance, DIA q-values). The deliverable matrix is the LAST place a fault becomes visible, so the agent inspects raw signal and removes contaminants BEFORE normalizing, because median normalization erases loading failures.

Prerequisites

pip install numpy pandas scipy scikit-learn matplotlib seaborn
# R packages: install.packages('PTXQC')   # PTXQC is on CRAN, NOT Bioconductor
# BiocManager::install(c('limma', 'MSstatsTMT'))

Quick Start

Tell your AI agent what you want to do: - "Plot raw per-sample boxplots and ID counts before I normalize, and flag loading failures" - "Strip MaxQuant contaminant and decoy rows before log-transform and normalization" - "Compute replicate correlation on log2 and CV on the linear scale, then run PCA colored by batch" - "Diagnose whether my missing values are MNAR or MCAR before I pick an imputer"

Example Prompts

Inspect Before Normalizing

"Show raw, un-normalized boxplots per sample with ID counts and total signal, and flag any sample shifted more than 2-3x below its group"

"Remove rows flagged Potential contaminant, Reverse, or Only identified by site from my MaxQuant proteinGroups before I normalize"

"What fraction of summed intensity is keratin and trypsin, and is it higher in my low-input samples?"

Reproducibility

"Calculate within-group Pearson correlation on log2 intensities and tell me if any sample correlates better with a different group"

"Compute median CV per condition on the linear scale, not on log data"

"Are my technical replicates above r 0.98 and is the biological CV in the 20-40% range?"

Missing Values

"Plot present-fraction versus abundance to decide if missingness is MNAR or MCAR"

"Filter to proteins valid in at least 70% of replicates in one condition before imputing"

"Which imputation method matches my missingness mechanism, and why would the wrong one corrupt my results?"

Batch and Outliers

"Run PCA on the normalized survivors and test whether PC1 or PC2 associates with batch rather than condition"

"Is batch the dominant axis of variance, and should I correct it before differential testing?"

TMT and DIA

"Check TMT channel-loading balance on raw reporter intensities and flag any channel deviating more than 2x"

"How many protein groups pass at 1% global q-value, and why is precursor q not enough?"

What the Agent Will Do

Load the un-normalized search output and intensity matrix
Plot raw per-sample boxplots, ID counts, total signal, and missing fraction; flag loading/injection failures
Filter contaminant, reverse, and only-identified-by-site rows before any transform
Log2-transform and normalize on the surviving good samples (mechanics route to quantification)
Compute replicate Pearson r on log2 and median CV on the linear scale
Diagnose the missingness mechanism and apply a completeness filter before imputing
Run PCA, test for batch association, and flag outlier samples
For TMT, inspect channel balance and isolation interference; for DIA, summarize global q-value counts

Tips

Inspect raw signal first; once the matrix is median-normalized the evidence of a loading failure is mathematically gone.
Almost no metric has a universal cutoff; trend each metric against your own per-instrument rolling baseline (Levey-Jennings, +/-2 SD warn, +/-3 SD action).
Read every metric with its co-readouts: a protein-count drop means spray, column, or sample depending on what co-moves.
Correlate on log2, never raw; a few abundant proteins make raw r meaningless.
Compute CV on linear intensity (or geometric CV on logs); always state normalization, transform, and software or the CV is uninterpretable.
Diagnose MNAR vs MCAR before choosing an imputer; the wrong one either kills or fabricates differences.
Document and justify every sample exclusion, and run a sensitivity check with and without borderline samples.

data-import - Load search-engine output and intensity matrices before QC
quantification - Normalization and imputation mechanics that QC mandates running AFTER inspection
differential-abundance - The moderated statistical test QC gates
dia-analysis - DIA q-value/FDR internals behind the protein-count QC
data-visualization/dimensionality-reduction-plots - PCA/MDS projection plotting
workflows/proteomics-pipeline - End-to-end pipeline placing QC before differential testing