Hi all,

I’m working on making RNA-seq reanalysis more reproducible and recently reproduced a published rice salt-alkali stress RNA-seq study starting from the raw FASTQ files.

Dataset: PRJNA895747

Study context: Rice seedlings under Na2CO3 salt-alkali stress, with untreated controls. The dataset has 12 paired-end RNA-seq samples across 4 groups with 3 biological replicates each.

Workflow used:

• adapter and quality trimming
• read alignment
• gene-level quantification
• differential expression with DESeq2
• PCA, volcano plots, heatmaps
• functional interpretation / enrichment-ready gene lists
• full provenance capture, including parameters, software versions, outputs, and generated code

The goal was to test whether an automated bioinformatics AI agent could reproduce the main biological signal from a published study without manual pipeline tuning.

Disclosure: I’m affiliated with Pipette.bio, which is the tool used for this run.

I’m looking for critique from people who run RNA-seq analyses regularly.

I’d especially appreciate feedback on:

1. Are these workflow choices reasonable for this dataset?
2. What QC checks would you require before trusting the differential expression results?
3. When reproducing an older RNA-seq study, how much difference in DEG counts would you consider acceptable if references, software versions, and filtering choices differ?
4. What would you want to see in a provenance report before considering an automated reanalysis credible?

Writeup: https://pipette.bio/blog/reproducing-rice-salt-stress.html

Use case report: https://pipette.bio/use-cases/rice-salt-stress.html

Original study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9840837/

Thanks for any feedback. I’m particularly interested in whether the workflow/provenance is enough for a reviewer or collaborator to evaluate the result without rerunning everything from scratch.