Hi,
I am analyzing the RNA-seq of a pilot project trying to detect markers in the blood of 5 treated patients.
The data I have is:
- 5 patients (no controls)
- 4 timepoints (1 pre-treatment + 3 post-treatment).
- 1 replicate per patient ---> 20 samples total.
I have already analyzed the differences between timepoints using DESeq2 (design = ~timepoint + patient). Now we are interested in finding individual patients/genes whose (temporal) response is different from the other patients. As we only have 1 sample for each patient x timepoint, it is not possible to run interaction models in DESeq2.
I was wondering if there is any method suited to detect genes where one patient's expression (either at one specific timepoint or the overall temporal pattern) is different from the others. To later (manually) find association with the patient's metadata.
I know the statistical power of such analysis would be crap, but... is there any method that can help with this, or should I make up something ad-hoc? Or should I just focus on genes significant on 2/3+ contrasts in the patient factor of my DESeq2 model?
Thanks
Without replicates or controls, I am not sure any of this data is really that useable. RNAseq is quite variable at the best of times, which is why you almost always use at least technical replicates if not biological replicates. Also, if you are looking to detect markers related to the treatment, why are you looking for genes that are different among patients?
But if you really just want to identify genes that are different among patients at each timepoint, this would be a simple R/Python script to write yourself using transcript counts.
This is a pilot analysis of a project with ~15 patients. Only 5 have been sequenced in this phase. Patients respond very differently to the treatment, with different prognoses and side effects/toxicity. That's why we are interested in finding some candidate genes showing a different expression profile through time in the currently sequenced patients. To later validate them in the larger sample using qPCR or further sequencing.
I have also run DESeq2 with a model using the polynomial of timepoints, to detect parabola-shaped genes. I'm wondering if there is some already established method that can detect the shape of the time-series for each gene x patient, instead of having to do that coarsely to check if a gene goes e.g., Up-Flat-Down, or Down-Flat-Flat.
The typical way to do this would be to group the samples by responsiveness, and compare the groups to each other.
I wish I could do that. However, these 5 samples were done in 2 batches. 2 males / 3 females. 3 tumor reduction / 1 NR / 1 tumor growth. Wildly different mean expression (across all timepoints) per patient of many genes (independent of batch)...
That's why I'm looking for a dirty method to get an idea of the "differences in shape" within each patient, and explore from there.
I could take each patient x gene, normalize the expression at each timepoint by t0, make a vector with c( t1-t0, t2-t1, t3-t1 ), calculate the pairwise distance between all patients. And then... what? Select the genes with the largest SD in their pairwise distances?