🪨 rosetta

Python interface to R/Bioconductor — pandas in, pandas out, .report() when you're done.

pip install rosetta-bioc

30-second demo

import rosetta as rb

# DESeq2 differential expression — one call, pandas out
results = rb.deseq2(counts_df, metadata_df, design="~ condition")
results.report()

DESeq2 Results Summary
──────────────────────────────
Total genes tested:      12,000
Significant (padj<0.05): 843 (7.0%)
  ↑ Upregulated:         428
  ↓ Downregulated:       415
LFC range:               [-4.71, 3.50]

That's it. No R code. No rpy2 boilerplate. No type conversion. Just results.

Three-Tier API

# Tier 1 — quick: one call, done
results = rb.quick_deseq2(counts_df, metadata_df, design="~ condition")

# Tier 2 — class-based: build up state, chain methods
seu = rb.Seurat(matrix).normalize().find_clusters().umap()

# Tier 3 — functional: explicit steps, full access
dds = rb.wrappers.deseq2.run_deseq2(counts, meta, design="~ batch + condition")
res = rb.wrappers.deseq2.get_results(dds, lfc_threshold=1.0)

Complete example — copy, paste, run

import pandas as pd
import numpy as np
from rosetta import deseq2

# Simulate RNA-seq counts: 1000 genes, 6 samples (3 control, 3 treated)
np.random.seed(42)
counts = pd.DataFrame(
    np.random.negative_binomial(5, 0.1, size=(1000, 6)),
    index=[f"gene_{i}" for i in range(1000)],
    columns=["ctrl_1", "ctrl_2", "ctrl_3", "treat_1", "treat_2", "treat_3"],
)

metadata = pd.DataFrame(
    {"condition": ["control"] * 3 + ["treated"] * 3},
    index=counts.columns,
)

results = deseq2(counts=counts, metadata=metadata, design="~ condition")
print(results.sort_values("padj").head(10))

Requires: Python 3.9+, R 4.0+, and Bioconductor's DESeq2 (BiocManager::install("DESeq2")).

What it wraps

All functions return a RosettaDataFrame (pandas DataFrame subclass) with a .report() method.

Not a toy — full design support

• Multi-factor designs: design="~ batch + condition", interaction terms, blocking factors
• LFC thresholds: proper hypothesis testing via lfcThreshold (not post-hoc filtering)
• Shrinkage: apeglm, ashr, normal — via lfc_shrink()
• Contrasts: contrast=["genotype", "mutant", "wildtype"]
• QC/normalization/outliers: DESeq2's size factors, Cook's distance, independent filtering all run normally — Rosetta doesn't hide the fitted object
• Weights, correlations: limma-voom with duplicateCorrelation, sample weights — everything the R function accepts, Rosetta passes through

Show me the R code

Don't trust a black box? Turn on codegen to see exactly what's running:

import rosetta as rb
rb.codegen.enable()

dds = rb.wrappers.deseq2.run_deseq2(counts, meta, design="~ batch + condition")
res = rb.wrappers.deseq2.get_results(dds, lfc_threshold=1.0)

  R> library(DESeq2)
  R> dds <- DESeqDataSetFromMatrix(countData=counts, colData=metadata, design=~ batch + condition)
  R> dds <- DESeq(dds)
  R> res <- results(dds, alpha=0.1, lfcThreshold=1.0)

rb.codegen.last() returns it as a string — paste into R to reproduce independently.

Modular DESeq2 API

For more control, use the step-by-step interface:

from rosetta.wrappers.deseq2 import run_deseq2, get_results, lfc_shrink

dds = run_deseq2(counts_df, metadata_df, design="~ condition")
res = get_results(dds, contrast=["condition", "treated", "control"], alpha=0.05)
shrunk = lfc_shrink(dds, coef="condition_treated_vs_control", type="apeglm")

res.report()
shrunk.report()

Enrichment analysis

import rosetta as rb

# Over-representation analysis
go_results = rb.enrich_go(gene_list, org_db="org.Hs.eg.db", ont="BP")
go_results.report()

# KEGG pathways
kegg = rb.enrich_kegg(gene_list, organism="hsa")
kegg.report()

Setup

Python side:

pip install rosetta-bioc

R side (one-time):

Rscript install.R

Or manually:

BiocManager::install(c("DESeq2", "edgeR", "limma", "clusterProfiler"))

Posit Cloud: See docs/posit-cloud.md for zero-config setup.

Requirements

• Python 3.9+
• R 4.0+ with Bioconductor
• rpy2 ≥ 3.5

Philosophy

1. Rosetta calls R — it doesn't reimplement it. All statistics run in the original, validated R packages.
2. Pandas in, pandas out. No R objects leak into your Python workflow.
3. Fail early, fail clearly. Input validation happens in Python before crossing the R boundary.
4. .report() everything. Results should be immediately interpretable without manual inspection.
5. Show your work. codegen prints the equivalent R code so you can verify, reproduce, or learn.

Contributing

See CONTRIBUTING.md. Good first issues are labeled — start with Issue #1: report() enhancements.

Contributors

• Catherine Chi Chung — GSoC 2026 contributor
• Matias Salibian Barrera — GSoC co-mentor, UBC Statistics

Acknowledgments

Built on rpy2 and the extraordinary R/Bioconductor ecosystem. All credit for the statistical methods goes to the original R package authors.

Supported by:

• Google Summer of Code 2026 — funding Catherine's development work
• JPMorgan Chase — startup banking and advisory through their Innovation Economy program
• AWS — quantum computing infrastructure via Amazon Braket
• Nodes Bio, Inc. — project lead, CI/hosting, and engineering

GSoC 2026 · MIT License · Nodes Bio