I'm doing some preliminary analysis on a list of genes of interest, and I was using a data set that contains RNA-seq data from patients and the corresponding clinical information, including survival. For a given gene, I split the patients up into two groups, "high" expression (>= median expression of the gene) and "low" expression (< median expression), plotted Kaplan-Meier curves, and did a log-rank test to see if survival was significantly different. I was surprised that none of the results were significant, even though those genes have been shown to divide patients into two groups with significantly different survival in other data sets.
I checked my data and realized that the genes were not normally distributed--I had thought they were already normalized, but I guess they weren't. My question is, though, would whether the data is normalized or not make a difference to the outcome of my analysis? I'm dividing the patients based on the median, so the patients in the upper 50% would be the same regardless of whether I normalized or not, and the patients in the lower 50% would be the same regardless of whether I normalized or not, since it's based on rank, not so much the actual values. Right? Or am I missing something?
Dear Lon,
1. Please make sure you are working with normalized data across samples before making any comparison.
2. Try comparing the 1st quartile (low-expressing) vs the 3rd quartile (high-expressing) instead of using the median.
3. If you are going to use a parametric test to compare your groups, applying some transformation (e.g. logarithm) in order to make the distribution normal is also recommended.
4. Adding 10E-04 to all genes it is also helpful when dividing by 0 in order to capture increase or decrease in gene expression.
Regards,
Rolando
Dear Lon,
1. Please make sure you are working with normalized data across samples before making any comparison.
2. Try comparing the 1st quartile (low-expressing) vs the 3rd quartile (high-expressing) instead of using the median.
3. If you are going to use a parametric test to compare your groups, applying some transformation (e.g. logarithm) in order to make the distribution normal is also recommended.
4. Adding 10E-04 to all genes it is also helpful when dividing by 0 in order to capture increase or decrease in gene expression.
Regards,
Rolando
In my experience plotting log2(normalized_rsem + 1) and dividing the samples by the mean value (high = above mean; low = below mean) is sufficient to capture significant differences. Although it may be useful to try different cutoffs, the statistical power decreases with decreasing sample size. It may be instructive then to plot the confidence intervals to visualize the extent of overlap between lines.
Dear Lon,
you could also try to run ROC curve analysis using survival status as classification variable and gene expression data as numerical variable to obtain potential cut-off value of gene expression that you can further test in survival analysis. This is not always optimal approach but might help at finding best cut-off.
Additionally, if you did your molecular analyses on both diseased and control patients, you can run ROC curve analysis using disease status as classification variable to find "optimal" cut-off value of gene expression for discriminating diseased and control subjects, then test whether this cut-off value discriminates survival among diseased patients (some kind of normal vs high expression among diseased). I'm not sure whether your data have to follow normal distribution for ROC curve analyses to be reliable or not, however this is only auxiliary first step and your cut-offs need to be tested on survival data as a next step and you can try to play around given cut-off value to see whether it affects prognosis or not. Data separated at median or at quartiles do not need to follow normal distribution.
If you don't obtain significant result, but see that survival curves diverge apart, this might be due to too small number of events in your data-set. Your study is then probably underpowered to detect statistical significance for this particular analysis.
I hope this would help you, Kind regards, Marko
Hi Lon-
Normalization is a precondition for analysis of Gene expression data.Therefore. one must do normalization of data.with best wishes.
ajit
thanks for recommending my answer@
- Syed Ismyl Mahmood Rizvi
- Yehya A. Salih
Dear Lon
RNA-seq data does not follow a normal distribution and even after normalization it is very difficult that you will get it. Indeed, the distribution of the reads counts follows a negative binomial distribution, in fact, most of the programs which are used to perform statistical analysis of RNA-seq data do not apply models based on normality. So my suggestion would be to apply non-parametric tests to check for differences in survival.
Kind regards
Riccardo
Bioinformatics data analyst at www.sequentiabiotech.com
- Badges
- Science topic
- Similar topics
- Biological Science
- Genetics
- Gene Expression