One of the driving questions in Alzheimer’s research has been its nature—is AD simply an extension of normal aging, or is it a disease unto itself? If it is a disease unto itself, what changes over time in the brain to make the aged tissue so much more vulnerable to attack? Using high-level bioinformatic approaches, Miller, Oldham, and Geschwind take a closer look at the interplay between aging and AD using transcriptional profiles from previously published data sets.

On a technical scale, this work is extremely thorough and careful. It is a terrific example of a well-reasoned meta-analysis in its truest form, paying attention to statistical probabilities at each stage of the analyses (e.g., probability of overlap between the two studies based on the discovery power in either). Rather than simply comparing lists of genes from two studies, the authors stripped the information from each study down to its most raw form (at least as raw as they could get it, i.e., CEL files from Affymetrix array scans), and used a consistent probe level algorithm across both studies to create...  Read more

One of the driving questions in Alzheimer’s research has been its nature—is AD simply an extension of normal aging, or is it a disease unto itself? If it is a disease unto itself, what changes over time in the brain to make the aged tissue so much more vulnerable to attack? Using high-level bioinformatic approaches, Miller, Oldham, and Geschwind take a closer look at the interplay between aging and AD using transcriptional profiles from previously published data sets.

On a technical scale, this work is extremely thorough and careful. It is a terrific example of a well-reasoned meta-analysis in its truest form, paying attention to statistical probabilities at each stage of the analyses (e.g., probability of overlap between the two studies based on the discovery power in either). Rather than simply comparing lists of genes from two studies, the authors stripped the information from each study down to its most raw form (at least as raw as they could get it, i.e., CEL files from Affymetrix array scans), and used a consistent probe level algorithm across both studies to create new data sets that are more comparable.

Using the innovative WGCNA approach to clustering, the authors established modules of genes based on similar behavior across disease states or aging. Within these modules, the authors established connectivity among genes and identified “hub” genes that appeared to be most often linked to other genes within the module. Among these, two genes stood out: VDAC1 in the mitochondrial module of AD, and YWHAZ, a fairly uncharacterized, but extremely abundant gene product in brain in a module that stubbornly refused to reveal a functional association. Further, well-known PSEN1 was found to be related to myelinating processes in a “guilt-by-association” module of myelin-related gene products. The authors found that two of the modules from the AD data set (synaptic and mitochondrial) were related to a single module from the aging study that contained both functional categories.

The implication here is that aging and AD do share some common processes. However, the question remains as to whether AD represents a frankly different process than aging. That’s because there are both intriguing agreements (discussed in the paper), and disagreements (for instance, that the mitochondrial and synaptic genes split into two distinct modules in AD, but resided in one aging module) between the two studies compared here.

It is remarkable that these relationships were seen even despite the confounds on a technical level of different labs, and times, as well as the biological confounds of different tissue type (hippocampal CA1 versus forebrain cortex). It would be interesting to know, in future studies using this approach, what the level of agreement would be between two studies attempting to examine the same disease state—such as AD—in different brain regions. Our own observations are that changes are more consistent at the functional group level as opposed to the per gene level.

Other interesting further work might include determining the consequences of various cutoff decisions in the course of implementing the algorithm, e.g., number of presence calls, variability for inclusion, lack of connectivity for exclusion, should number of connections for determining hub genes be adjusted for number of genes in module, etc. Also interesting would be to determine “anti-modules,” functional groups that remain static across groups.

View all comments by Eric Blalock