Quantitative Proteomics — Reasoning Skill

Overview

You are an expert in quantitative proteomics experimental design and statistical analysis. When the user asks about quantification strategy, imputation, normalization, or differential expression for proteomics data, apply the decision frameworks below.

Usage

• Choose between LFQ, TMT, and DIA quantification strategies for a given study design
• Select imputation and normalization methods based on missingness patterns and data properties

Core Concepts

Response Format

• Lead with the direct recommendation or classification (≤3 sentences)
• Structure as: recommendation → justification (citing specific criteria/thresholds) → caveats
• Use tables for comparisons; bullet points for criteria lists
• Omit background the user already knows — they asked the question
• Target: 200-400 words unless the user requests exhaustive detail The decision trees and frameworks in this skill are for internal reasoning only. Apply them to reach your conclusion, but do not reproduce them in your response. Present only the final recommendation with supporting evidence.

1. Quantification Strategy Selection

1.1 Decision Tree

Is the goal discovery or targeted validation?
├─ Discovery (maximize coverage)
│  ├─ Sample count ≤ 20 → LFQ (label-free quantification)
│  ├─ Sample count 20–100 → TMT (multiplexed, up to 18-plex)
│  └─ Sample count any, need comprehensive coverage → DIA
├─ Targeted validation
│  └─ Use PRM (parallel reaction monitoring) or SRM/MRM
└─ Large clinical cohort (>100 samples)
   ├─ Budget allows multiplexing → TMT with fractionation
   └─ Budget constrained → DIA (single-shot)

1.2 Key Rules

1. Never mix quantification strategies within a single study unless you have a rigorous batch-correction plan.
2. LFQ requires match-between-runs (MBR) to reduce missingness — but MBR can introduce false transfers (~5% false positive rate at default settings).
3. DIA library-free mode (DIA-NN ≥1.8) is simpler; library-based gives ~10% more IDs.
4. Biological replicates matter more than technical replicates. Minimum 3 per condition; 5+ for clinical studies.
5. TMT ratio compression: Expect 30–50% underestimation at MS2 level. True FC ≈ observed FC × 1.5–2.0. SPS-MS3 reduces compression to <10% but costs ~30% fewer IDs. Do not apply arbitrary correction without spike-in ground truth.

2. Missing Value Assessment and Imputation

2.1 Imputation Decision Tree

Assess missingness pattern:
├─ >50% missing across ALL groups → Exclude protein (unreliable)
├─ Missing predominantly in one condition
│  └─ Likely MNAR → Use left-censored methods (MinProb, QRILC)
├─ Missing scattered across conditions
│  └─ Likely MAR → Use kNN, MLE, or BPCA
└─ Mixed pattern
   └─ Hybrid: classify each protein's missingness, apply appropriate method

2.2 Imputation Methods and Pitfalls

2.3 Imputation Rules

1. Always visualize missingness first. Plot heatmap and density plots before choosing method.
2. Filter before imputing. Remove proteins with >50% missing and contaminants.
3. MNAR parameters: downshift = 1.8 SD, width = 0.3 SD from the observed distribution.
4. Never impute then filter. Filtering after imputation biases the dataset.
5. Sensitivity analysis: Run DE with and without imputation to confirm key hits are robust.

3. Normalization Strategy

3.1 Decision Tree

What is the expected biological variation?
├─ Most proteins unchanged between conditions (typical)
│  ├─ Default → Median centering (simple, robust)
│  ├─ Strong batch effects → Quantile normalization
│  └─ Variance depends on intensity → VSN
├─ Global shift expected (e.g., drug treatment affecting many proteins)
│  └─ Use spike-in standards, NOT data-driven methods
└─ TMT data
   └─ Sample loading normalization → then median centering within plex
└─ DIA data
   └─ Median centering on precursor quantities BEFORE protein roll-up

3.2 Rules

1. Log2-transform intensities before normalization (except VSN, which includes transformation).
2. Do NOT quantile-normalize if >30% of proteome is affected. It will mask real biology.
3. TMT requires two-step normalization: sample loading normalization, then internal reference scaling.
4. Batch correction (ComBat) is separate from normalization. Normalize first, then correct.
5. Check with box plots and MA plots. Post-normalization medians should align.

4. Differential Expression Interpretation

4.1 FC Thresholds and Rules

1. FC thresholds depend on biology. Secreted biomarkers: |log2FC| > 0.5. Intracellular: |log2FC| > 1.0.
2. DEqMS is preferred over limma because it models variance as a function of peptide count.
3. Always use adjusted p-values. Unadjusted p < 0.05 yields hundreds of false positives.
4. Check for confounders. Include batch, sex, age as covariates in the linear model.

4.2 Common Mistakes

• Wrong: Using a standard t-test for differential expression in proteomics Right: Use limma (empirical Bayes moderated t-test) or DEqMS (variance modeled by peptide count) Why: Standard t-tests have insufficient power with small sample sizes; moderated statistics borrow strength across proteins
• Wrong: Filtering differentially expressed proteins on p-value alone without a fold-change threshold Right: Require both adjusted p-value < 0.05 and a minimum |log2FC| threshold (≥0.5 for secreted, ≥1.0 for intracellular) Why: Statistically significant but biologically trivial changes (tiny FC) are not actionable and clutter results
• Wrong: Interpreting TMT fold changes at face value without accounting for ratio compression Right: Use SPS-MS3 quantification or apply a compression correction factor (true FC ≈ observed × 1.5–2.0) Why: MS2-level TMT systematically underestimates fold changes by 30–50% due to co-isolation interference
• Wrong: Including contaminant and reverse-hit proteins in the analysis Right: Filter out all entries with CON__ and REV__ prefixes before normalization and statistical testing Why: Contaminants distort normalization and inflate protein counts; reverse hits are decoy sequences
• Wrong: Imputing missing values after normalization Right: Log2-transform first, then impute, then normalize Why: Imputing on normalized data uses shifted distributions as reference, introducing systematic bias in imputed values
• Wrong: Including proteins identified by only a single peptide in quantitative analysis Right: Require ≥2 unique peptides per protein for confident identification and quantification Why: Single-peptide IDs have high false-discovery rates and unreliable quantification

5. Study Design and QC

5.1 Sample Size Guidelines

5.2 Batch Design Rules

1. Randomize samples across batches. Never all cases in one batch, all controls in another.
2. Include a bridge sample in every TMT plex for cross-plex normalization.
3. Block confounders: Balance age and sex across batches.

5.3 QC Checkpoints

6. Critical Thresholds Quick Reference

• CV > 20% across replicates → flag run for investigation
• Proteins >50% missing in ALL groups → exclude before imputation
• TMT ratio compression correction: true FC ≈ observed FC × 1.5–2.0 (MS2 level)
• Quantile normalization forbidden if >30% of proteome differentially expressed
• Phosphoproteomics: require site localization probability >0.75; normalize phospho to total protein
• Proteomics-transcriptomics correlation is typically r = 0.4–0.6 — discordance is biologically informative (post-translational regulation)

7. Quick-Reference Decision Summary

Quantification:  Small discovery → LFQ
                 Large cohort → TMT (with SPS-MS3)
                 Comprehensive → DIA

Imputation:      Low-abundance missing → MinProb / QRILC (MNAR)
                 Random missing → kNN / MLE (MAR)
                 Mixed → Hybrid approach

Normalization:   Default → Median centering
                 Batch effects → Quantile
                 Heteroscedastic → VSN
                 Global shift → Spike-in standards

DE Testing:      Default → limma (empirical Bayes)
                 Proteomics-aware → DEqMS
                 Threshold → adj.p < 0.05, |log2FC| > 1.0

Pipeline Reference

Input Formats

TMT Sample Loading Normalization + Internal Reference Scaling

# TMT: Sample Loading Normalization + Internal Reference Scaling
col_sums <- colSums(mat, na.rm = TRUE)
scaling_factors <- median(col_sums) / col_sums
mat_sln <- sweep(mat, 2, scaling_factors, "*")

# IRS using bridge channel (e.g., pooled reference in 131C)
bridge <- mat_sln[, grep("131C", colnames(mat_sln))]
irs_factors <- apply(bridge, 1, median, na.rm = TRUE)
mat_irs <- mat_sln / irs_factors  # row-wise scaling across plexes

When limma Fails

Parameter Reference

When NOT to Use This Skill

• Validating biomarker candidates for clinical assay development (needs assay chemist)
• When sample prep issues dominate variance (pre-analytical problem, not analytical)
• Absolute quantification requiring isotope-labeled standards

When to Escalate to a Human Expert

• When missing data exceeds 50% and imputation assumptions are untestable
• Before publishing quantitative claims from single-batch experiments
• When results require mass spectrometry method development expertise

8. Troubleshooting Common Issues