Pick a score that fits your design

Within each experimental setting, you still have some selection of scores to pick from. Results might be quite different. What are the problems? To

How many conditions?

Samples are dependent/paired/ somehow connected?

Select score:

Select cutoff a:

Samples are dependent/paired/ somehow connected?

f-score SAM, Limma Wilcoxon Correlation pAUC

f-score SAM, Limma Wilcoxon Correlation

F-score ^ time-SAM, Limma series

Kruskal-Wallis ^ survival Correlation analysis f-score SAM, Limma Wilcoxon Correlation

Control or estimate significance: FWER = a or FDR = a

Figure 18.1.

The roadmap to the gene list.

start with, 'differentially expressed' is not a well-defined term. Figure 18.2 displays gene expression values of two genes A and B in two conditions (light and dark gray). The y-axis shows expression levels on additive scale, and two horizontal lines indicate the mean expression in each condition. The distance between the two lines is larger for gene A than it is for gene B. Is gene A the better candidate? Gene B is rather constant in both groups, while gene A seems to be quite variable. Should we take variance into account? We can do so, using t-scores. The t-score is higher for gene B. Now, B is the better candidate.

You reach a different result with a third argument. For gene B, the expression values in the two conditions share only a small overlap. For gene A, this is not the case. Note that two dark gray values exceed the light gray mean. Do you feel that this is an important observation? You can formalize it using the pAUC-score. It is higher for gene B, consistent with the t-score but inconsistent with the fold change. By reflecting on Figure 18.2, we came up with three different perspectives on differential gene expression, and they led to conflicting results. Applied to a microarray experiment, the three scores result in different rankings of genes. Eventually, the rankings can be so diverse that you will be left with bad feelings when it comes to biological conclusions. The problem is what is a good ranking? We are not able to answer this question. In Section 18.3 we will discuss several scores and point out pitfalls. The final choice is upon

0 0

Post a comment