Senior Staff Scientist
At a Glance
Name: Gary Churchill
Title: Senior Staff Scientist, the Jackson Laboratory
Professional Background: 2003 — present, senior staff scientist, the Jackson Laboratory, Bar Harbor, ME; 1998 — 2003, staff scientist, Jackson Laboratory; 1990 — 1996, assistant professor, biometrics unit, Cornell University.
Education: 1988 — PhD in biostatistics, University of Washington, Seattle; 1983 — Bsc in mathematics, Massachusetts Institute of Technology.
WASHINGTON, DC — Even though array industry attendance was sparse at Cambridge Healthtech Institute's Total Microarray Data Analysis and Interpretation conference, held here last week, many exhibitors proved that microarray technology is a way to explore microorganisms in space and to create single-experiment tests for a variety of airborne viruses.
One researcher claims he can use three dyes to compare differentiation in gene expression profiles, as compared to the more popular two-dye and one-dye methods. The study, recently published by researchers from the Jackson Laboratory in Bar Harbor, Maine, and the University of Connecticut in a supplementary edition of Bioinformatics, the study, entitled "Experimental design for three-color and four-color gene expression microarrays," explores how using more than two dyes in an experiment can "increase design efficiency and power to detect differential expression without additional samples and arrays."
The switch to using more dyes in an experiment could increase the utility of microarrays with "no additional expenses for most facilities that are already equipped with two-color microarray instruments," the authors argue in the paper.
To learn more about what kind of impact this could have on the industry, BioArray News spoke with paper co-author Gary Churchill from the Jackson Lab this week.
So what was the premise for this project?
Well, there are basically one-color arrays and two-color arrays, and I must say that the one-color arrays are really dominating the market. We use them now, they work great. One reason they work great is because of precision manufacturing. They are really precise. Homemade arrays tend to be really sloppy.
The two-color system is just ingenious. And I don't think the people who invented it realize how ingenious it is. The basic idea behind it is to pair two measurements on the same substrate. That's an old idea called blocking. It's an idea from the 1920s and 30s and agricultural statistics where, if you wanted to compare things that are grown on a heterogeneous background, you pair them together in groups called blocks, and the blocks in agriculture were actually plots of land, and the land might vary in moisture and fertility and so on. But by putting things together in blocks and measuring several blocks you could make valid relative comparisons. So that relative comparison idea comes up.
So in microarrays the two colors provide you the ability to measure two different RNAs on the same microarray, so the block slide is two. But the bigger the block size — the more efficient the experiment. So that's the basic idea. We thought, well, if a block size of two is a great idea, and it is, then a block size of three is even better, and four and five and so on. Of course there is a bit of a diminishing return, the improvement of three over two is much better that four and five and so on.
So it was just an idea, and the other part of it is this. The two dyes, if you look at the way they behave, over their range of intensities, the gain function is non-linear. And it's not identical for the two dyes. And this has created a whole furor of research and normalization, the most popular method being the low S method for kind of smoothing out the non-linearity in the dye response. But something we noticed a long time ago is that dye swap fixes everything.
The dye swap design is where you take two arrays and you compare two samples, one red to green, and one green to red. And if you work out the algebra, with arbitrary gene functions, everything cancels out and you get these absolutely beautiful, robust estimates.
But why is there a need for dye swapping in the first place?
Well, you know the red and green dyes don't behave the same, so if you measure one array and you measure samples red to green, you have to do something. You have to fix the fact that red and green aren't the same. And it's not that they are different from a proportional constant, which would be easy to fix.
What actually happens is that if you draw the graph they crisscross. At one intensity range the dyes behave one way, but at another intensity range the dyes behave another way. So there's this intensity-dependent dye effect, which is a real nuisance, and frankly it must be there in the Affy arrays, but, you never know. So the idea of dye swapping is to compare your samples and then to reverse the direction of the dyes, compare them again. And what that does is with one sample, everything is biased in the red to green direction, in the other measurement in the other array, everything is biased in exactly the opposite way. So the biases cancel each other out. So the idea of dye swapping is extremely powerful.
But if you put the ideas together, blocking is a key idea and dye swapping is a key idea, so what if we had three dyes? You could do better experiments if you could figure out what the three dye equivalent of a dye swap is. And that's what Yan Woo did — I posed the question and he quickly came up with the cyclic design. So to do a three color dye swap, we have three colors — red, yellow, green, I guess — you compare your three samples — red to green, green to yellow, yellow to red, in a little triangle.
But how did you get in contact with the UConn group in this research?
Well, we had this theoretical result that sat around in our drawer for a year, and somehow Winfried Krueger [from UConn Medical School's Laboratory for Microarray Technology] came up to visit for some reason and we got to talking to him, and he said that he was busy trying to develop a three-color system in his lab and so we talked him into doing experiments. What we learned is that in an ideal world, the three dye system should be better than the two-dye system. In the real world, the dyes don't behave the way you want them to. In fact he used three different Alexa dyes, and one of them was a problem.
And in the balance I think when we actually empirically compared the results, the two-dye and the three-dye results came out neck and neck. What I want to add to that is that the amount of optimization that people have put into two-dye systems is amazing, and as far as we know, Winfred has tinkered around with the three-dye system and it has a lot of growth potential if we can just figure out the technical details of implementing it. It is interesting to know that you can do better using more dyes, that there are some technical challenges that kind of prevent you from achieving the theoretical optimum, but that could be worked on.
Why three dyes? Why not four or five?
The gain from two to three is huge, and the gain from three to four is modest and so on, so the amount that you gain from adding another dye adds diminishing returns. So we thought that just going from two to three would be the biggest pay off in practical terms. And given that there were issues with the dyes and practicalities, I am sure that ideally you could use 100 dyes, but I am sure it would become too technically complicated.
But then why hasn't anybody done this yet commercially?
Why do people not use more dyes? Well I think again the dominance of Affymetrix has occurred because they have really focused on reducing the variability and minimizing that variability and I think they've also effectively ignored some things, like non-linear dye response, which drives me nuts that they can just blatantly ignore it, but they get away with it and the damn things work. They work. What can I say? I use them. And they are cheap. And we've made our own spotted arrays, and we've made them for five or six years now. And although there are labs that can produce them cheaper, it's a long painful process, and if you add up all the labor and the cost it adds up to more money than just going out and buying a commercial array.
There is a company called NimbleGen, and they have two color arrays built on the similar design as Affy, and I know they do two-color experiments, and we at one time tried to talk them into doing three color experiments, and I think they just didn't get it, and we didn't really push too hard to persuade them. Maybe they'll read our paper and come back.
Is it such a leap of faith in the industry for people to think that if they added another dye they'd get better results?
Well, again there's this trade off between theory and reality. In theory you absolutely do — you get better results in three. But then there's the daunting task of defining the three color assay, which properties do they have, how do we scan them, how do we make the arrays, what are all the protocols for getting the dyes on. People are appropriately daunted by the task. It's not a simple thing to debug a new technology like that.
What would be your advice to someone that was interested in doing some work on this?
I would say, to use our designs and focus on improving the technology. I think what we've come up with are optical experimental designs. I would absolutely use the three color version of the dye swap. Failure to balance the dyes across your experiments is always going to lead to misleading results. And always dye swap, it fixes all the non-linear junk that normalization is supposed to fix, but doesn't.