Hi everyone, I’m analyzing a bacterial RNA-seq experiment with two conditions: TR (treated): 5 biological replicates and CTL (control): 5 biological replicates

What I observe (problem): In unsupervised QC (PCA on VST/rlog, sample distances), two CTL controls consistently appear “intermediate” between CTL and TR: CTL-4 and CTL-5.

HISAT2 overall alignment rate: 
CTL-1/2/3: ~97%, CTL-4: 91.90%
CTL-5: 91.34%  
TR samples: ~93.9–96.2% 

featureCounts assignment is in the ~30–41% range for all samples; CTL-4/5 are slightly lower than other CTL but not dramatically. Expression-based diagnostics supporting “TR-like CTL”.

I computed an “TR marker score” using a set of TR marker genes (mean VST expression across the marker set):

TR samples: 10.11–10.43 
CTL-1/2/3: 3.55–3.71 
CTL-4: 7.8687 
CTL-5: 7.8482 

I also computed a centroid-correlation diagnostic (top 2000 variable genes; correlation to CTL centroid vs TR centroid): For CTL samples, delta = cor_to_CTL - cor_to_TR:

CTL-1/2/3: ~0.322–0.326 
CTL-4: 0.0405 
CTL-5: 0.0502 

So CTL-4/CTL-5 correlate almost equally to CTL and TR centroids (much closer to TR than the other CTL replicates).

Differential expression approaches compared I compared 4 scenarios (FDR < 0.05 and |logFC| > 1):

All samples (5 TR vs 5 CTL) - limma-trend on logCPM 
DE genes: 990 

All samples (5 TR vs 5 CTL) - limma::voomWithQualityWeights() 
DE genes: 1527 

Clean CTL only (5 TR vs 3 CTL) dropping CTL-4/CTL-5 - limma-trend 
DE genes: 1636 

Clean CTL only (5 TR vs 3 CTL) dropping CTL-4/CTL-5 - voomWithQualityWeights 
DE genes: 1612 

Sample quality weights (scenario 5 TR vs 5 CTL with voomWithQualityWeights): 
CTL-4: 0.186 
CTL-5: 0.192 
other CTL: 2.287–2.878 (also TR-4 and TR-5 have lower weights: 0.441 and 0.570) 

So voomWithQualityWeights strongly down-weights CTL-4/CTL-5.

Sensitivity / stability checks: When I drop both CTL-4 and CTL-5 and compare DE results to the baseline voomQW-all analysis:

baseline voomQW (5 vs 5): nDE = 1527 
drop CTL-4/CTL-5 voomQW (5 vs 3): nDE = 1612 

common DE genes: 1510 
same direction: 100% 
correlation of logFC: 0.9998 

Set overlap (Jaccard) on DE gene sets:

voomQW (all) vs standard (clean): 
Jaccard = 0.836 (common=1440) 

standard (all) vs standard (clean): 
Jaccard = 0.559 (common=942) 

standard (clean) vs voomQW (clean): 
Jaccard = 0.851 (common=1493) 

Lab records indicate no sample mix-up/swapping (consistent sample tracking during extraction/library prep). Given CTL-4/CTL-5 look biologically TR-like by marker score and centroid correlation (not just noisy), what’s the most defensible way to proceed?

1. Main analysis = 5 vs 5 with voomWithQualityWeights (and report weights)?
2. Main analysis = 5 vs 3 (drop CTL-4/CTL-5) with voomWithQualityWeights, and report 5 vs 5 voomQW as sensitivity?
3. Should I treat CTL-4/CTL-5 as a third group (“suspected mislabel / contamination / partial treatment”) and exclude them from the primary contrast?

In your experience, which additional checks are most convincing to distinguish mislabel/contamination vs batch/technical artifact when I don’t have genotyping/fingerprinting? Thank you very much to anyone who is willing to help me.