jgmeyer/mississippi

HepG2 Drug Perturbation Proteomics Dataset

This dataset contains a high-throughput drug perturbation proteomics screen performed in HepG2 cells. The raw quantitative data are provided in AnnData .h5ad format. Sample-level annotations are stored in adata.obs, and drug-level annotations are provided separately in the accompanying drug information Excel file.

This dataset accompanies the preprint:

Robotic perturbation proteomics and AI agents enable scalable drug mechanism discovery
Yuming Jiang, Cameron S. Movassaghi, Jesús Muñoz-Estrada, Niveda Sundararaman, Amanda Momenzadeh, Jesse G. Meyer
bioRxiv DOI: 10.64898/2026.05.04.722718

Dataset summary

The study profiled HepG2 cells treated with an FDA-approved drug library using a semi-automated proteomics workflow, rapid LC-MS/MS acquisition, and downstream AI-assisted interpretation. The screen includes drug-treated samples, DMSO vehicle controls, deferoxamine positive controls, and QC samples across multiple experimental batches.

Key characteristics reported for the study:

• Cell model: HepG2
• Treatment duration: 24 hours
• Drug concentration: 10 µM
• Proteomics acquisition: LC-MS/MS on an Orbitrap Astral operated in DIA mode
• Scale: 1,232 proteomes
• Protein features: 8,703 quantified proteins
• Drug library: 172 compounds, with six biological replicates per compound in the original design
• Controls per plate: QC samples, DMSO vehicle controls, and DFO positive controls

Files

Expected files in this repository:

File	Description
raw.h5ad	Raw proteomics matrix in AnnData format. Samples are stored as observations and protein measurements are stored as variables/features.
drug_details.xlsx	Drug-level metadata, including compound annotations.
README.md	This dataset description.

Data format

The primary data file is an AnnData object in .h5ad format.

Typical loading example:

import scanpy as sc

adata = sc.read_h5ad("your_dataset_file.h5ad")
print(adata)
print(adata.obs.head())

The matrix in adata.X contains the proteomics abundance matrix. Depending on the exact uploaded file version, values may represent raw or transformed protein abundance values. Users should inspect preprocessing metadata and value distributions before downstream analysis.

Sample metadata: adata.obs

Sample-level metadata are stored in adata.obs. The expected structure is:

Column	Description	Example values
drug	Compound or sample identifier. QC samples are labeled as QC_sample; controls and drug treatments use their corresponding names.	QC_sample, DMSO, Deferoxamine, Loratadine, Methylene blue
batch	Experimental or acquisition batch/plate identifier.	Blue1, Blue2, Blue3, Blue4, Blue5, Blue6, Red1, Red2, Red3, Red4, Red5, Red6, RedMix
sample_type	Broad sample class.	Drug, DMSO, DFO, QC

Example rows from adata.obs include QC samples and drug-treated samples from the same batch:

index	drug	batch	sample_type
Blue2_A3	QC_sample	Blue2	QC
Blue2_A4	QC_sample	Blue2	QC
Blue2_B10	Triclabendazole	Blue2	Drug
Blue2_B11	Camphor	Blue2	Drug
Blue2_B12	Fexofenadine (hydrochloride)	Blue2	Drug

Drug metadata

The accompanying drug information Excel sheet should be treated as compound-level metadata. It can be joined to adata.obs using the drug or compound name field. This metadata is useful for interpreting perturbation profiles by known compound class, target, pathway, indication, or other annotations included in the spreadsheet.

Users should verify exact column names in the Excel sheet before joining. For example:

import pandas as pd

metadata = pd.read_excel("drug_details.xlsx")
print(metadata.columns)

Batch effects and recommended preprocessing

This dataset contains visible batch structure and should not be analyzed as if all samples were directly exchangeable without correction. PCA visualizations of the raw data show strong batch-associated structure, including separation of batches and QC samples. After ComBat correction, the batch-associated structure is reduced, but users should still evaluate whether correction is appropriate for their specific analysis.

Recommended preprocessing considerations:

1. Inspect missingness at both the sample and protein levels.
2. Filter low-quality samples and proteins with excessive missingness. I suggest removing samples with less than 7,600 non-zero values.
3. Log-transform abundance values if this has not already been done.
4. Impute missing values using a strategy appropriate for proteomics data.
5. Correct batch effects using a method such as ComBat, while preserving biological groups of interest.
6. Re-check PCA/UMAP after correction, colored by both batch and sample_type.
7. Use corrected data for differential abundance, clustering, and mechanism-of-action analyses, unless the goal is specifically to evaluate raw technical variation.

A conservative batch-correction example:

# Example only. Confirm value scale and metadata before use.
import scanpy as sc
import pandas as pd
from combat.pycombat import pycombat

# adata: samples x proteins
expr = pd.DataFrame(
    adata.X,
    index=adata.obs_names,
    columns=adata.var_names,
)

# pycombat expects features x samples
corrected = pycombat(expr.T, batch=adata.obs["batch"]).T

adata.layers["combat_corrected"] = corrected.loc[adata.obs_names, adata.var_names].to_numpy()

Important: batch correction can remove real biological signal if batch is confounded with treatment, sample type, or plate layout. Before applying correction, inspect the design matrix and confirm that biological contrasts of interest are represented across batches.

Suggested analyses

Potential uses of this dataset include:

• Drug versus DMSO differential protein abundance analysis
• Replicate-level drug response consistency analysis
• Mechanism-of-action inference from proteomic signatures
• Drug-drug similarity analysis using whole-proteome response vectors
• Pathway enrichment analysis using ranked protein-level fold changes
• Evaluation of batch correction methods for high-throughput proteomics
• Benchmarking AI or agentic workflows for biological interpretation of perturbation proteomics data

Differential abundance analysis

For drug-level analysis, compare each drug against DMSO controls after appropriate filtering, transformation, missing-value handling, and batch correction. The associated manuscript used Welch's t-test for each drug versus DMSO and adjusted p-values within each drug using the Benjamini-Hochberg procedure, with proteins considered significantly regulated using an FDR threshold and log2 fold-change cutoff.

Example structure:

# Pseudocode only
for drug in sorted(adata.obs.loc[adata.obs["sample_type"] == "Drug", "drug"].unique()):
    drug_samples = adata.obs["drug"] == drug
    dmso_samples = adata.obs["sample_type"] == "DMSO"
    # compare corrected abundance values between drug_samples and dmso_samples

Caveats

• The h5ad file should be treated as the source of the raw sample-level proteomics matrix.
• Batch effects are visible in the raw data and should be evaluated carefully.
• The drug information Excel file is required for compound-level interpretation beyond the adata.obs fields.
• QC samples are technical controls and should generally not be mixed with drug-treated samples for biological differential abundance analysis.
• Observational or mechanistic claims require follow-up validation; the dataset is best used as a perturbation resource for hypothesis generation and prioritization.

Citation

Please cite the associated preprint if using this dataset:

@article{jiang2026robotic,
  title = {Robotic perturbation proteomics and AI agents enable scalable drug mechanism discovery},
  author = {Jiang, Yuming and Movassaghi, Cameron S. and Muñoz-Estrada, Jesús and Sundararaman, Niveda and Momenzadeh, Amanda and Meyer, Jesse G.},
  year = {2026},
  journal = {bioRxiv},
  doi = {10.64898/2026.05.04.722718}
}

Contact

For questions about the dataset, contact the corresponding authors listed in the associated preprint.