In your recent Nature Genetics paper, you used cDNA microarrays from populations of killfish species, Fundulus heteroclitus and Fundulus grandis, to look at variations in gene expression within particular populations and between populations. What did you find?
About 18 percent [of genes] were significantly different between genetically similar populations — they were all healthy and there were no adaptive differences between those individuals in a population. So, 18 percent of 10,000 [genes] you see [expressed] in humans should be different between two healthy individuals. That’s actually the most significant thing about that paper. What you expect when you compare individuals, say for cardiac problems or for cancer, is that 18 percent of the genes will be different from one another — a statistically significant difference — even though they have nothing to do with the phenotypes you are looking at.
How do you know that finding could be extrapolated from fundulus to humans?
Because there is probably more variation in humans, because humans have a larger population size. What we do know is the nucleotide variation in fundulus, and in fact in most vertebrates, is very similar to what you find in humans. There is about an 0.1 percent difference between any two humans: that means about three million base pairs between you and I, or you and anybody, on average. And that number is actually greater than what you find in fundulus. So, based upon about how much variation you have in DNA sequences out there, that variation in DNA sequences is what affects or causes the change in gene expression.
In your experiment, you used a loop design from Gary Churchill’s group at the Jackson Labs, in which each array contains a combination of two among 15 samples, and each sample ends up being compared to four others in the loop. Also Churchill uses the ANOVA, or analysis of variance, where within-group and between-group variation is compared. Why did you use this loop design, rather than a reference RNA design, in which each sample is compared to a reference?
The loop design is just a balanced experimental design that allows you a more powerful statistical analysis. If you just look at a fold difference, which most microarray papers do, there could be a huge difference in hair color between you and I, and that difference represents no more than just random noise. My hair could be gray, but that grayness has nothing to do with the genomic phenotype. That kind of variation, both physiological, within and between individuals, is something you can explore with analysis of variance. Now, for microarrays, there’s also an additional variance, and that’s the experimental variation. If I measured how much genes you have [expressed] today and tomorrow, I could find a two-fold difference in my favorite gene between you today and you tomorrow, just because of the experimental variation. The trouble is, when you don’t do replication, and you don’t have a loop design, there is no way to know whether that two-fold or greater difference is due to experimental variation or just genetic variation. The loop design was done specifically so we could get a handle on how much experimental variation there was, and how much variation there was between individuals, and therefore how much variation between populations.
So with a reference design you would have had to compare every single replicate to a reference...
A reference design doesn’t give you as many replications for the same number of slides. If you have three individuals, with a loop design you would get six measurements because there would be two individuals on each [slide.] With a reference design, you would get three measurements of the reference, and one measure of each individual. And that doesn’t give you any statistical powers. You could do more replication, but then you have this problem: Imagine a numerator and a denominator, and both have some variance. Let’s say the denominator varies from 1 to 4, and the numerator varies from 4 to 8. In some experiments, [the numerator is] 4 and [the denominator] is 1 [so there is] a fourfold difference, and in others, [both numerator and denominator are 4], and there is no difference. So a ratio of two experimental numbers, the reference and the experimental number will have a greater variance, because you have variance below and variance above, and together you will have much greater variance. That’s what’s wrong with the reference design, and that’s why we didn’t use it. The bottom line is that a loop design is statistically much more powerful, and minimizes experimental and measured variance.
Did you actually work with Gary Churchill on this?
My postdoc, Marjorie Oleksiak, went up there and worked in his lab for a couple of weeks, getting his Matlab programs (for ANOVA) down. (See http://www.jax.org/ research/churchill/software/anova/index.html)
What do your findings about within-group variation imply for other researchers using gene expression data?
Lets take the example of a breast cancer [gene expression profiling study.] I think they took 15 individuals, and they broke down five different types of breast cancer. There were 8,000 genes in that array. Basically, it comes down to 50 genes between all five different groups. What my data says is that there is only 18 percent difference between individuals. Well, 18 percent is greater than 50 some-odd genes they found on those arrays. So the point is, if you are going to use microarrays to look at variation between human individuals for disease states, you must first ask how much variation you find in normal individuals. In other words, if you think gene X is important for cancer because it is twice as great in the cancer patient [than the normal one], is it also twice as variable between non-cancer individuals? There is variation in expression among individuals. You have to take that variation into consideration before you start assigning genes as being important.