nodalpoint.org

Ok, so with a slight time lag, here are my contributions. Rather predictably, they revolve around microarrays. Specifically, about the much vaunted power of microarrays when applied to molecular medicine, with particular reference to cancer profiling.

Michiels et al (Lancet editorial by Ioannides, Nature news story) reanalyse data from seven large cancer expression profiling studies, and conclude that, overall, predictive gene sets do not replicate across data sets.

Now, it's tempting to conclude that microarrays are therefore useless, but I think that's wrong. The problem is fourfold, and lies with us, not the technology:

The last two points in particular I think are key: there seems to be a prevalent attitude that doing one array and using package X on default will magically give you the cure to cancer.

There is one great weakness in this meta-analysis: these authors do not appear to use the original raw data. Instead, they take processed (normalised) data from public repositories, and proceed with their analyses. What is not clear from this report is the differences in residual systematic signal in each (sub)data set - which are likely to be considerable, given that many rudimentary normalisation procedures fail to remove major bias trends.

So we are left with an interesting question: are Michiels et al correct in stating that the results across studies replicate poorly, or do they commit the same basic error? IMNSHO, I think their results are broadly right - the findings of microarray studies don't correlate greatly (See Miklos and Maleszka for another example). This however, does not mean that the problems are intractable with this technology. It just means you have to try harder - and use raw data.