On a given microarray design there are multiple different probes spotted for many genes. The (normalized) signals of the features (all referring to the same gene) often are quite different (log2 values can vary between 2 and 16, so essentially from "almost undetectable" to "completely saturated").
If a gene set analysis or an over-representation analysis is performed, there should be one value per gene.
How to select which signal to use for the gene? I don't feel good to take the average of all the multiple features, because they are often so different. Taking the highest signal only also seems to be wrong.
Any ideas?
The attached file shows a table with example data (from an Agilent Microarray) with 5 different probes addressing the gene "PRDM". The last 4 colums show the log signal intensities for 4 different samples. The values range from 3 to 10, so there is a more than 100-fold difference in the signal intensities between the probes.
I hope I correctly understood your question, but my suggestion is use standardization of multiple features then, take the average of all the multiple features.
Thanks, Maxim.
R is not the problem. The problem is the rationale behind what to select or average and how. The gene sets (KEGG, GO, etc) do not refer to any transcript variants, as far as I know. To see if a set is enriched or over-represented, one somehow needs "one value per gene", whene "gene" is whatever is annotated in the respective gene set collection.
The choice to select those probes that give signals with the highest variation among the experimental conditions sounds interesting. I would expect that this probe will likely capture that one variant that at least is somehow differentially regulated. And since the gene set annotation does not distinguish different variants, such sets will get lower p-values (a better "ranking") when there is evidence that at least particular variants of the participating genes are differentially expressed (whether or not these variants do play a role in the particular set or pathway).
I wonder if there is a paper about that topic.
Using the mean (or median) may have the opposite effect, and using the probe with the maximum may predominantly select probes with a large amount of unspecific signal (I have seen one example, where one probe had a signal of 16, where all others (for the same gene) had signals around 3. The probe sequence that gave the high signal had a long stretch of GpT so I think that the signal "collects" everything from mRNAs containing some CpA streches.
Again I wonder if there is a paper about that topic.
As Micheal stated: is is likely that the method won't matter much (except for small sets I suppose), but I think that chosing the probes with highest variation is the most reasonable w.r.t. the aim of getting ideas which gene sets might be critically affected.
I use the probes that show the maximum average signal. This set of probes shows the most resonable pattern in a volcanoplot. If e.g. by average the signal from different probes (per gene), the pattern in volcanoplots look "strange".
https://www.networkanalyst.ca/. NetworkAnalyst software may help you.
Can I use following criterion to select probes and corresponding expression values in case these probes show same gene annotation? Jochen Wilhelm Michael B Black
Monther Alhamdoosh :1st priority "_at"
2nd priority "_a_at"
3rd priority "_s_at"
4th priority "_x_at"
This is based on their description here: https://www.affymetrix.com/support/help/faqs/mouse_430/faq_8.affx
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning