sigRecon Benchmarking

This repository contains the benchmarking pipeline and figure-generation code for the sigRecon manuscript. It evaluates signature recontextualization methods — algorithms that take a gene signature from one biological context (cell line, tissue) and predict the corresponding signature in another context.

The reusable R package, including method implementations and evaluation utilities, lives at montilab/sigrecon.

Overview

Methods are benchmarked across four datasets under three data-availability regimes:

Regime	Description
Control (null)	No target-context perturbations used
Low coverage (1/10)	1/10th of target perturbations available
High coverage (9/10)	9/10th of target perturbations available

Performance is measured by Jaccard similarity and fgsea NES between predicted and ground-truth target signatures, reported as improvement over the source baseline.

Repository Structure

sigrecon/          R package (methods + evaluation functions)
scripts/           Per-dataset pipeline scripts (02–08), numbered by dataset
  02_PerturbSeq/
  03_DrugMatrix/
  04_SciPlex/
  06_Tahoe/
results/
  eval/            Pre-computed evaluation tables (.rds) per dataset/method/regime
  figures.Rmd      Master figure-generation notebook (reads from results/eval/)
drug_metadata.parquet  Drug annotation table used across scripts

Datasets

Folder	Dataset
02_PerturbSeq	Perturb-seq (K562, RPE1)
03_DrugMatrix	DrugMatrix (kidney, liver)
04_SciPlex	SciPlex (K562, MCF7)
06_Tahoe	Tahoe (multiple cell lines)

Methods Benchmarked

Method	Description
Mean	Context mean as predicted signature
NetProp	Network propagation via WGCNA co-expression graph (netProp())
projCor-Eigen / projCor-GSVA	Projection-based scoring then gene reranking (projectCor())
Orthos	Neural network–based context transfer
scGPT	Foundation model–based signature prediction
Stack	Stacked regression baseline

Reproducing Figures

# From repo root
rmarkdown::render("results/figures.Rmd")

Requires the MLAB environment variable pointing to the upstream data directory. Pre-computed evaluation tables in results/eval/ are sufficient to regenerate all paper figures without re-running the full pipeline.

Data

Full-size pseudobulk expression and perturbational signatures for each dataset are archived on Zenodo:

Dataset	Perturbational signatures	Pseudobulk expression
DrugMatrix	10.5281/zenodo.21432933	10.5281/zenodo.21433031
SciPlex	10.5281/zenodo.21432935	10.5281/zenodo.21433011
Perturb-seq	10.5281/zenodo.21432937	10.5281/zenodo.21433138
Tahoe	10.5281/zenodo.21433000	10.5281/zenodo.21433050