Back to Case Studies

Case study · Perturbational proteomics

HepG2 Drug-Perturbation Proteomics Across 172 Compounds

A public HepG2 DIA-MS dataset was audited for missingness and batch structure before batch-aware differential-abundance modeling. Drug signatures and pathway enrichments are presented as response patterns and hypotheses, with RedMix confounding made explicit.

Homo sapiens1,232 samples172 drugs8,630 proteinslimma
View full Pipette session Run a similar analysis

Original 35-page session report · PDF · 3.7 MB

At a glance

Large-scale proteomic response mapping with a visible batch boundary

The dataset covers 172 compound perturbations, but drug identity and acquisition series are not fully separable. The workflow distinguishes modeled contrasts, visualization-only batch correction, and exploratory mechanism hypotheses.

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

Which proteomic response programs remain visible after accounting for technical structure?

The analysis compared each compound with DMSO, clustered continuous drug-response signatures, and summarized Hallmark pathway enrichment. Mechanism-of-action labels remain hypotheses rather than confirmed targets.

Which drugs produce broad or shared protein-abundance changes in HepG2 cells, and which apparent effects cannot be separated from batch or control structure?
Sourcejgmeyer/mississippi
AssayHepG2 drug-perturbation DIA-MS proteomics
Samples1,232 total; 1,052 drug and DMSO samples in differential testing
Controls23 DMSO, 24 DFO, and 155 pooled QC samples
Features8,630 measured proteins; 7,559 tested in limma contrasts
02 / Analysis workflow

Missingness audit, batch-aware limma models, signatures, and enrichment

Inspect and transformNon-negative intensities were log₂ transformed; zeros were treated as missing.
Audit quality and batchDetection rates, missingness, PCA, acquisition batches, and the RedMix control structure were examined before testing.
Fit abundance modelslimma modeled drug and batch for 160 non-RedMix compounds using BH FDR < 0.05 as the sole significance criterion.
Build signaturesPer-drug moderated t-statistics were correlated and hierarchically clustered across 7,559 proteins.
Interpret cautiouslyHallmark enrichment used continuous ranks; removeBatchEffect was used only for a visualization matrix, never for differential testing.
Review correctionDownstream rankings and figures were regenerated after removing an unplanned fold-change filter. FDR < 0.05 is the documented criterion for the reported significant-protein counts.
03 / Results

Broad response programs coexist with strong technical and biological confounding

Cell-cycle suppression

E2F_TARGETS and G2M_CHECKPOINT were downregulated across several signature clusters.

Proteome collapse

Eight non-RedMix drugs showed reproducibly low protein detection across most or all replicates.

RedMix boundary

Twelve RedMix-only drugs lack an in-batch DMSO control, so drug and series effects cannot be separated.

Exploratory biomarkers

Pan-drug stress candidates reflect broad response and are not validated drug-specific markers.

PCA of HepG2 proteomics samples colored by acquisition batch
Figure 1. PCA by acquisition batch. Batch explains R² = 0.59 of PC1 in the report, making technical structure a primary modeling constraint.
Top HepG2 drugs ranked by FDR-significant protein count
Figure 2. Perturbation-strength ranking. RedMix-only compounds are flagged because their apparent effects are confounded with the acquisition series.
Hallmark pathway enrichment summary across HepG2 drug-response clusters
Figure 3. Hallmark pathway summary. Cluster-average enrichment shows recurring cell-cycle suppression; pathway and mechanism labels remain exploratory.
04 / Limitations

Batch, missingness, and design constrain mechanism claims

  • Batch explains 59% of PC1 in the main PCA and drug identity is tied to acquisition series.
  • The 12 RedMix-only compounds have no in-batch DMSO control, so their drug effects are not identifiable.
  • Overall missingness is 10.6%; 758 proteins are missing in more than half of samples.
  • The report does not provide a complete sensitivity analysis for imputation or missing-not-at-random behavior.
  • Broad loss of detected proteins may reflect cytotoxicity or global proteome disruption rather than target-specific action.
  • Signature similarity, pathway enrichment, and proposed biomarkers require external validation.
Evidence boundary

The non-RedMix contrasts support batch-aware response estimates. The analysis does not validate drug targets, causal mechanisms, clinical efficacy, or RedMix-only drug effects.

05 / Reproducibility and outputs

Artifacts recorded in the Pipette session

drug_vs_control_DE_results.csvPer-drug limma results
drug_perturbation_ranking.csvFDR-based response ranking
drug_signature_matrix_tstat.csvContinuous drug signatures
pathway_enrichment_per_cluster.csvExploratory Hallmark results
analysis_flags.jsonMachine-readable limitations
limma_fit_objects.rdsFitted statistical models
Original Pipette session report35-page PDF · 3.7 MB · generated July 6, 2026. The HTML article is the primary case study; the PDF is the preserved session record.
View session Download PDF
Analysis systems
AI Agent vs Workflow Manager
Pathway analysis
GSEA vs GO Enrichment

More drug-discovery case studies

Imatinib Docking Against ABL1 (PDB 2HYY) →

Analyze a perturbational proteomics dataset

Bring an abundance matrix and sample metadata, including batches and controls, to make the comparison design auditable.

Start in Pipette