PlantPerturb-Small

PlantPerturb-Small is a reproducible small-data benchmark for predicting the transcriptional response of Arabidopsis thaliana ecotypes to mild drought from their baseline gene-expression profiles.

Research draft

The current manuscript is available as paper/PlantPerturb-Small-Manuscript-Draft.pdf. Its editable Quarto source and BibTeX references are included beside it.

This is a research draft. Author information, funding, and competing-interest statements remain placeholders, and the model comparisons require repeated nested cross-validation before publication claims are finalized.

Main result

Model	Test MSE	Test MAE	Pearson	Spearman
Zero change	0.1451	0.2729	undefined	undefined
Development mean	0.1366	0.2666	0.2155	0.1705
PCA–ridge	0.1302	0.2578	0.2518	0.2142
Nearest neighbors	0.1413	0.2711	0.1939	0.1646
PLS	0.1300	0.2587	0.2429	0.2056

PCA–ridge provides the strongest balance across metrics. PLS has a 0.2% lower mean squared error, which is too small to establish superiority with 14 held-out ecotypes. On the independent E-MTAB-3279 study, learned models had higher error than zero-change prediction, showing substantial study distribution shift.

Data

• Primary study: E-MTAB-5009
• External study: E-MTAB-3279
• 89 complete ecotypes in the primary analysis
• 17,633 retained transcripts after CPM filtering
• 62/13/14 ecotype-level train/validation/test split using seed 42

The public raw RNA-seq files are not duplicated in this repository. The included download and preprocessing scripts reconstruct the analysis from the accessioned studies.

Repository contents

• paper/: manuscript PDF, Quarto source, and references
• model/: fitted PCA–ridge model and machine-readable metadata
• scripts/: data preparation, training, and evaluation programs
• results/: compact metric tables and publication figures
• notes/: step-by-step research notes

Reproduce the manuscript

Install Python, Quarto, and Typst, reconstruct the processed data with the included scripts, then run:

quarto render paper/plantperturb_manuscript.qmd --to typst

Intended use and limitations

This repository is intended for research, teaching, and method development. It is not a validated predictor for agricultural decisions. Most ecotype-condition combinations lack independent biological replicates, bulk leaf RNA-seq averages different cell states, and the external evaluation shows that model calibration does not transfer reliably between studies.