We thank Jan Hultgren for his important and stimulating comments on our paper, and would like to offer the following responses:

1. I am afraid we do not fully understand the objection here. Underlying the M-values are of course data on the relative expression level of each spot from each individual. For every individual (4 per treatment group), one microarray (plus technical replicates) were run. The M-value, by definition, estimates the difference in average expression between two treatments or groups of animals. The statistical analysis developed for microarray analysis is based on the variation between experimental units (biological replicates), and we therefore continue to believe that the analysis we have done is the most appropriate one.

2. Hultgren is correct in stating that the spots on one micro-array are inter-dependent, since regulatory networks will affect many genes in the same way. However, the main point of the correlation analysis is that there is a dependency of M-values (which, as already said, estimate the average differential expression) between generations, something which is not biologically obvious. The important message here is that this dependency was only seen in the population of birds where we simultaneously saw a phenotypic effect.

3. The randomization procedure used for calculating an average coefficient of correlation does of course not render independent datasets, as Hultgren rightly points out. However, we think this was the best available type of analysis for our purpose. We could have simply calculated a correlation coefficient based on all 9033 pairs of M-values from the two generations, but this would be a biased analysis with an inflated p-value due to the facts that there were so many data-points, and that most of them had small absolute M-values. We could also have decided to look at only the most differentially expressed genes (those shown in the fig), but again this would have rendered a biased r. We therefore used a randomization procedure and picked random subsets of M-values as described in the paper. Even if the randomly generated datasets are not independent, they do provide a clear indication that the M-values of the two generations are correlated. However, we agree that the SD and the actual p-value of the t-test must be treated with care.

4. As we do say in the paper, there was no overall effect of treatment on the average (or median) number of tests needed to reach the learning criterium in the offspring, which is probably why Hultgren does not find any significant result in his analysis. However, by visual inspection, it is obvious that the learning curves appear quite different between Leghorn offspring of stressed and control birds. After 5,6 and 7 rounds of testing, the differences between treatment groups were visually striking, and we therefore decided to test the differences in proportion of birds having solved the task at the points which appeared different. We did not apply a correction for multiple testing, because we only tested three data-points in the offspring, and in the parental generation, the differences were already shown by an ANOVA. We admit that the effects are small in the offspring, but still clearly discernible.

Cheers,
Per Jensen, on behalf on the authors.