Analysing the same data in several ways can show how sensitive a result is to the analysis. Testing whether it holds in another experiment needs a separate comparison. That was what I looked for when revisiting a Paeonia–Glycyrrhiza transcriptome reanalysis. On 3 September 2026, I decided not to extend the same analysis without an independent, comparable source of evidence.
Transcriptomic data use RNA measurements to describe activity across many genes. This project reanalysed public data on how preparation methods for Paeonia and Glycyrrhiza related to gene-activity patterns in cells. A difference within the analysed data did not establish that the same interpretation would hold in another cell system or experimental setting.
More repository records did not mean more independent experiments
I examined where the additional data came from and what each experiment compared, rather than simply counting accession numbers. A repository such as GEO can split one research programme across several records. Separate files or accession numbers do not by themselves provide independent validation.
The required contrast also had to be present. Data comparing a preparation with a control cannot automatically test a difference between two preparation methods. Herb ratio, solvent, exposure concentration, cell system and experimental batch have to be recoverable well enough to distinguish preparation from those other conditions.
I classified the additional material against those requirements, looking for an experiment separate from the existing analysis that still contained the relevant comparison. The 3 September check did not establish that combination, so I stopped the rework and resubmission. Changing the analysis or expanding the explanation would not supply the missing comparison.
The public computational record remains
On 13 July, I had already released curated data, code, manifests and verification records in a Zenodo reproducibility package. It includes analyses of preparation-label separation within the data and prediction tests that leave a cell context out of the analysis before testing it. To follow the computation, a reader needs to see its inputs and the conditions being compared.
That package provides a route through the reanalysis; it is not independent experimental validation of its conclusions. It does not establish clinical efficacy, safety or patient benefit. Stopping the subsequent extension did not erase the released work.
Revisiting the question would need a separate source with the relevant comparison, and a rule fixed before examining outcomes for deciding whether the original interpretation survives. What I wanted to retain was more useful than a stopped-status label: an account of checking which evidential gap the additional data could actually fill.