I have a set of gene expression samples which are split in different subgroups. I would like to have a representative sample for each class. Is it correct and biologically significant to take an average (mean or median) sample (i.e. a class centroid)?

Is this methodology exposed to bias or is it safe to assume that the centroid of the class may be a good estimate of the expression value inside the class? Are there alternatives?


migrated from biology.stackexchange.com Nov 17 '16 at 18:27

This question came from our site for biology researchers, academics, and students.

  • $\begingroup$ You may be able to get a better answer by asking on Cross Validated, as this is fundamentally a statistics problem, not a biological one. $\endgroup$ – MattDMo Nov 17 '16 at 14:32
  • $\begingroup$ Hi @MattDMo, while I agree with you that this may be asked on CV, my focus in this question is that centroids may be of biological interest (not only from the stats point-of-view). This in the context of gene expression. If this does not fit the requirements I will forward it to CV. $\endgroup$ – gc5 Nov 17 '16 at 14:35
  • 1
    $\begingroup$ I would suspect that CV has some biology people there and get biology-related statistics questions all the time, so, while I don't usually recommend this, I'd suggest posting there as well (you don't need to delete this post in case you get a useful answer here) just to broaden the number of eyeballs seeing your post. $\endgroup$ – MattDMo Nov 17 '16 at 18:18
  • $\begingroup$ Ok, I will flag the post to a moderator to be exported to CV. Thanks $\endgroup$ – gc5 Nov 17 '16 at 18:24

This methodology definitely throws away parts of the information. However, the effect of this depends on what exactly you want to do with the representative samples, and on how the original data is distributed (how dispersed it is etc.)

  • $\begingroup$ I agree with you that some information is lost. I want to get the median sample for each class (i.e. centroid) to use them as references for correlation (e.g. Pearson's r) with unclassified samples. In this way I would like to correlate an unclassified sample with the representative sample of each class, and hopefully identify which class the sample is most correlated with. $\endgroup$ – gc5 Nov 17 '16 at 15:09
  • $\begingroup$ In that case it might be better to use all the data, and run a statistical test for significance of correlation with each entire class. Or use the evaluation methods that are common in your field (in good papers). $\endgroup$ – root Nov 17 '16 at 16:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.