---
name: mims-harvard/tooluniverse-proteomics-analysis
source: https://app.decimal.ai/s/mims-harvard-tooluniverse-proteomics-analysis@1/SKILL.md
source_sha256: 0cf7e2b2e830
---

# Proteomics Analysis

## RULE ZERO — Check for pre-computed results FIRST

Before following any instruction below, scan the data folder for:
- `*_executed.ipynb` → read with `tu run read_executed_notebook '{"data_folder":"<path>","search":"<keyword>"}'` and cite its cell outputs as the authoritative answer
- Pre-computed result files (CSV/TSV with names like `*results*`, `*deseq*`, `*enrich*`, `*stats*`, `*_simplified.csv`) → read directly and report the requested value
- Canonical analysis scripts (`analysis.R`, `run_*.py`, `find_*.R`, `*.Rmd`) → execute as-is and read the output

Only follow this skill's re-analysis recipe below if **none** of the above exist. Re-running from raw data produces different numbers than the published answer and is much slower (often 5-10× turn count).

---

Comprehensive analysis of mass spectrometry-based proteomics data from protein identification through quantification, differential expression, post-translational modifications, and systems-level interpretation.

## When to Use This Skill

**Triggers**: User has proteomics MS output files, asks about protein abundance/expression, differential protein expression, PTM analysis, protein-RNA correlation, multi-omics integration involving proteomics, protein complex/interaction analysis, or proteomics biomarker discovery.

## COMPUTE, DON'T DESCRIBE
When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.

## Core Capabilities

- **Data Import**: MaxQuant, Spectronaut, DIA-NN, Proteome Discoverer, FragPipe outputs
- **Quality Control**: Missing value analysis, intensity distributions, sample clustering
- **Normalization**: Median, quantile, TMM, VSN — choice depends on experimental design (see Interpretation Framework)
- **Imputation**: MinProb (MNAR), KNN (MAR), QRILC for missing values
- **Differential Expression**: Limma, DEP, MSstats for statistical testing
- **PTM Analysis**: Phospho-site localization, PTM enrichment, kinase prediction
- **Protein-RNA Integration**: Correlation analysis, translation efficiency
- **Pathway Enrichment**: Over-representation and GSEA for protein sets
- **PPI Analysis**: Protein complex detection, interaction networks via STRING/IntAct

## Workflow Overview

```
Input: MS Proteomics Data
    |
Phase 1: Data Import & QC
Phase 2: Preprocessing (filter, impute, normalize)
Phase 3: Differential Expression Analysis
Phase 4: PTM Analysis (if applicable)
Phase 5: Functional Enrichment (GO, KEGG, Reactome)
Phase 6: Protein-Protein Interactions (STRING networks)
Phase 7: Multi-Omics Integration (optional, protein-RNA correlation)
Phase 8: Generate Report
```

See [PHASE_DETAILS.md](PHASE_DETAILS.md) for detailed procedures per phase.

## Integration with ToolUniverse

| Skill | Used For | Phase |
|-------|----------|-------|
| `tooluniverse-gene-enrichment` | Pathway enrichment | Phase 5 |
| `tooluniverse-protein-interactions` | PPI networks | Phase 6 |
| `tooluniverse-rnaseq-deseq2` | RNA-seq for integration | Phase 7 |
| `tooluniverse-multi-omics-integration` | Cross-omics analysis | Phase 7 |
| `tooluniverse-target-research` | Protein annotation | Phase 8 |

## Quantified Minimums

- At least 500 proteins quantified (human: 3,000+ is reasonable; 10,000+ is deep coverage)
- At least 3 biological replicates per condition (non-negotiable for reliable statistics)
- Filter to proteins with 2+ unique peptides (single-peptide IDs are not reported as DE)
- Statistical test: limma or t-test with Benjamini-Hochberg multiple testing correction
- Pathway enrichment: at least one method (GO, KEGG, or Reactome)
- Report must include: QC summary, DE results with volcano plot, pathway analysis, visualizations

## Interpretation Framework

### Starting Point: Experimental Design

Quantitative proteomics compares protein abundance. LOOK UP DON'T GUESS — always verify the experimental method, platform, and replicate count before choosing an analysis strategy.

**Quantification strategy decision tree:**
- Cell culture, high accuracy needed → **SILAC** (ratios within same MS run, most accurate, but culture-only)
- Multiple conditions, multiplexing needed → **TMT/iTRAQ** (up to 18-plex in one run; TMM/VSN normalization; beware ratio compression artifact that reduces observed fold-changes)
- Discovery study, flexible design → **Label-free (LFQ)** (intensity-based; median/quantile normalization; more missing values; wider dynamic range)
- **Replicates**: n < 3 = unreliable fold changes (variance cannot be estimated). Minimum n = 3 biological replicates; n >= 4 preferred for clinical samples. Never report significance from duplicates.
- **FDR cutoff**: Benjamini-Hochberg correction mandatory. FDR < 0.05 standard; FDR < 0.01 stringent. Never report unadjusted p-values alone.

### Protein Identification Reasoning

Protein identification from MS data follows a logical chain. LOOK UP DON'T GUESS — search UniProt and STRING for protein annotation rather than inferring function from name alone.

1. **Peptide mass fingerprinting (PMF)**: Intact protein digested → measured peptide masses compared against theoretical digest of all database proteins. A match requires >=4 peptides covering >=15% of the protein sequence. Single-peptide hits are unreliable (could match multiple proteins).
2. **Tandem MS (MS/MS)**: Fragment ion spectra matched to peptide sequences via search engines (Andromeda, SEQUEST, X!Tandem). Each peptide-spectrum match (PSM) scored; only PSMs above FDR threshold count. Unique peptides (mapping to one protein) are essential — shared peptides cannot distinguish between protein isoforms.
3. **Protein inference**: Multiple peptides → protein groups. When peptides are shared between homologs, report the protein group (not individual proteins). Use `proteins_api_search` or `UniProt_search` to resolve ambiguous protein groups.
4. **Coverage matters**: 2+ unique peptides is the minimum for a confident protein ID. Proteins identified by a single unique peptide should be flagged as tentative.

### Post-Translational Modification (PTM) Analysis Reasoning

PTMs (phosphorylation, ubiquitination, acetylation, glycosylation) add biological complexity beyond protein abundance.

1. **Site localization**: A phospho-site is confidently localized only if the localization probability > 0.75 (MaxQuant) or ptmRS score > 75 (Proteome Discoverer). Ambiguous sites should not be reported as specific residue modifications.
2. **Enrichment is required**: Without phospho-enrichment (TiO2, IMAC), only the most abundant phosphopeptides are detected (~1% of phosphoproteome). An absence of a phospho-site in non-enriched data does not mean it is absent biologically.
3. **Kinase prediction**: If phospho-sites are identified, predict upstream kinases using motif analysis. Cross-reference with `OpenTargets_get_target_safety_profile_by_ensemblID` for kinase-disease associations. LOOK UP kinase-substrate relationships in PhosphoSitePlus rather than guessing from sequence motif alone.
4. **Stoichiometry**: A protein can be 5% or 95% phosphorylated at a given site — this matters enormously for function but is hard to measure. Report whether data supports stoichiometry estimation or only site identification.

### Differential Expression Thresholds

- **Strong**: padj < 0.01, FC > 2.0, ≥5 unique peptides, <20% missing
- **Moderate**: padj 0.01-0.05, FC 1.5-2.0, 2-5 peptides, 20-50% missing
- **Weak/unreliable**: padj 0.05-0.1, FC 1.2-1.5, 1-2 peptides (single-peptide proteins are unreliable), >50% missing (imputation needed)

### Evidence Grading

- **T1**: Validated by orthogonal method (Western blot, PRM) + functional study
- **T2**: Significant DE (padj < 0.05, FC > 1.5) in 2+ biological replicates
- **T3**: Significant DE in 1 experiment, or significant but low FC
- **T4**: Identified but not quantified, or single peptide identification

### Synthesis Questions

1. **How many proteins are differentially expressed?** (>500 DE proteins suggests global perturbation; <50 suggests targeted effect)
2. **Are key pathway proteins concordantly regulated?** (all subunits of a complex changing = high confidence)
3. **Do proteomics results correlate with transcriptomics?** (low correlation is common — post-translational regulation)
4. **Are PTM changes driving the phenotype?** (check phosphoproteomics if available)
5. **What is the coverage relative to the expected proteome?** (human: ~10K quantified is good; <3K is limited)

---

## Limitations

- **Platform-specific**: Optimized for MS-based proteomics (not Western blot quantification)
- **Missing values**: High missing rate (>50% per protein) limits statistical power
- **PTM analysis**: Requires enrichment protocols for comprehensive PTM profiling
- **Absolute quantification**: Relative abundance only (unless TMT/SILAC used)
- **Protein isoforms**: Typically collapsed to gene level
- **Dynamic range**: MS has limited dynamic range vs mRNA sequencing

## References

**Methods**: MaxQuant (doi:10.1038/nbt.1511), Limma for proteomics (doi:10.1093/nar/gkv007), DEP workflow (doi:10.1038/nprot.2018.107)

**Databases**: [STRING](https://string-db.org), [PhosphoSitePlus](https://www.phosphosite.org), [CORUM](https://mips.helmholtz-muenchen.de/corum)

## Reference Files

- [PHASE_DETAILS.md](PHASE_DETAILS.md) - Detailed procedures for each analysis phase, including report template