Back to Case Studies
Human HepG2 DIA-MS 172 drugs limma Batch-aware

HepG2 Drug-Perturbation Proteomics Across 172 Compounds

A first-party account of a large DIA-MS analysis: dataset inspection, missingness and batch assessment, batch-aware differential abundance, drug-signature clustering, exploratory pathway analysis, methodology-review corrections, and explicit confidence boundaries.

Samples1,232
Proteins8,630
Drugs172
Batches13
01 / Scientific question

Which drug-response programs are visible after accounting for the dataset’s technical structure?

The analysis examined a public HepG2 DIA-MS perturbation dataset to compare each compound with DMSO, cluster continuous proteomic signatures, and prioritize response programs. The report treats mechanism-of-action interpretations as hypotheses rather than confirmed targets.

Source dataset
jgmeyer/mississippi on Hugging Face
Assay
HepG2 drug-perturbation DIA-MS proteomics
Controls
23 DMSO, 24 DFO, and 155 pooled QC samples
DE matrix
1,052 drug and control samples after one placeholder exclusion
02 / Managed execution

Workflow and methodology-review corrections

The session preserved the distinction between statistical modeling and visualization-only correction, then regenerated downstream rankings after review.

Inspect and transformRaw non-negative intensities were converted to log₂ abundance; zeros were treated as missing.
Quantify quality and batchMissingness, detection rates, PCA outliers, acquisition batches, and the RedMix control structure were audited before modeling.
Fit differential-abundance modelslimma modeled drug and batch for 160 non-RedMix compounds using BH FDR < 0.05 as the sole significance criterion.
Build response signaturesPer-drug moderated t-statistics were correlated and hierarchically clustered across 7,559 tested proteins.
Interpret with boundariesHallmark enrichment used continuous ranks; a separate removeBatchEffect matrix was created only for PCA visualization, never for testing.
03 / Results

The strongest conclusions concern shared response programs and confounding—not individual drug mechanisms

Cell-cycle suppression

E2F and G2M programs were repeatedly downregulated across multiple drug-signature clusters.

Proteome collapse

Eight non-RedMix drugs showed broad, replicate-consistent loss of detected proteins consistent with cytotoxicity or global stress.

RedMix confounding

Twelve compounds lacked in-batch DMSO controls, preventing their apparent drug effects from being separated from batch or series.

Candidate biomarkers

The strongest candidates were predominantly pan-drug stress signals, not validated drug-specific biomarkers.

Top 25 HepG2 drug perturbations ranked by FDR-significant proteins, with RedMix-confounded drugs highlighted in red
Report figure: top perturbation rankings under the corrected FDR-only criterion. Red bars identify RedMix compounds without an in-batch control and must not be interpreted as validated top drug effects.
04 / Interpretation boundary

Limitations materially change what can be claimed

  • Batch explained R² = 0.59 of PC1 variance and was the dominant global technical axis.
  • The 12 RedMix compounds have no in-batch control, so their differential-abundance, ranking, and mechanism claims are lower-confidence.
  • Overall missingness was 10.6%; 758 of 8,630 proteins were missing in more than half of samples, and no alternative-imputation sensitivity analysis was performed.
  • Eight drugs showed broad proteome collapse, making targeted pathway interpretations unsafe without viability data.
  • Pathway enrichment, biomarkers, and mechanism hypotheses are exploratory and lack external validation by targeted MS, transcriptomics, or viability assays.
Confidence boundary

Moderate for the QC/batch characterization and recurring cell-cycle-suppression pattern; low-to-moderate for specific mechanisms; low for conclusions involving RedMix-only compounds.

05 / Provenance

Reproducibility artifacts recorded in the session

The report inventories the result tables, model objects, QC flags, corrected rankings, figures, and scripts used across the analysis and review cycle.

drug_vs_control_DE_results.csvPer-drug limma results
drug_perturbation_ranking.csvCorrected FDR-only perturbation ranking
drug_signature_matrix_tstat.csvContinuous response signatures
pathway_enrichment_per_cluster.csvExploratory Hallmark enrichment
analysis_flags.jsonMachine-readable interpretation caveats
limma_fit_objects.rdsFitted statistical model objects
Original session report35-page PDF containing the 11-section scientific report, methods, limitations, output inventory, figures, and methodology-review correction record.
Download PDF

Run a similar proteomics analysis

Bring your abundance matrix and experimental design to a managed workspace with explicit QC and review gates.

Start an analysis