Bulk RNA-seq: when batch was perfectly confounded with treatment
A startup's lead program rested on 2,800 differentially expressed genes. The first PCA component was sequencing run, and the run was the treatment.
- Client
- Seed-stage biotech
- Data
- Bulk RNA-seq, 12 samples, two sequencing runs
- Timeline
- 6 working days
- Outcome
- Avoided a false lead; balanced study confirmed 3 of 5 pathways
Anonymized composite case. Details are combined and altered to protect client confidentiality.
The problem
Treated and control samples were extracted on different days, prepared with different kit lots and sequenced on separate runs. The DE results looked strong, and the team was preparing an investor update around them.
What we found
- Batch and treatment were perfectly confounded, so no statistical method could separate them.
- PC1 (41% of variance) separated runs. Many top genes were known library-prep-sensitive genes (GC-rich and short transcripts).
- RNA integrity also differed systematically between the groups.
How we rescued it
- Quantified which signals were consistent with batch artifacts using GC content, transcript length and RIN-associated gene sets.
- Used RUVSeq with empirical negative-control genes to estimate how robust each pathway-level signal was. We were explicit that this could not fully resolve confounding.
- Designed a small, balanced follow-up of 4 vs 4 samples, randomized across extraction days and one sequencing run, with a power estimate from the existing data.
The result
The balanced study confirmed 3 of the 5 pathways the team cared about and refuted 2. The team presented an accurate, smaller story, and avoided building a program around a batch artifact.
| Measure | Result |
|---|---|
| Original DE genes | 2,800 |
| Pathways confirmed in balanced study | 3 of 5 |
| Cost of follow-up | A fraction of a failed program |