Morning Session, 3 September (11th MGED Meeting, 1-4 September, 2008)
Microarray data is analyzed with a pipeline of algorithms to remove non-biological sources of variability. 133, 056 analysis methods in bioconductor. Which steps of the pipeline are important? Which alorithms work the best? What are the interactions, if any? Consider all possible analysis methods. Evaluate perfomance vs. gold-standard. Use a linear modeling approach: calculate a metric (e.g. AUC), fit an ANOVA, calculate percent variance explained (PVE). The PVE tells you how much of the variability is due to that step or set of variables you're considering.
Which steps of the pipeline matter? Only two steps accounted for 93% of the variability: WN (within-array normalization) and DF (differential testing). Do algorithms interact? The default assumption is that there has been no interactions. used 2nd-order ANOVA. There are large interactions: WN and BN (between-array normalization) greatly antagonize each other.
They had 3 different test datasets. The results are the averages of the performances across those 3 datasets. 90% of the variability was accounted for by your choice of differential testing (DF) at the end. The rest of it was mostly to do with BG. RankProd algorithm performed best. Large interactions, as before: one example was NM (normalization method) and method for perfect-match mismatch adjustment (PM).
Which steps are important? Not all of them. Differential expression seems to be the driving factor (DF). Which worked best? Surprising result with minimal dataset dependence. Are there interactions among steps? Yes, so new algorithms should be evaluated "in context" instead of a static pipeline.
These are just my notes and are not guaranteed to be correct.
Please feel free to let me know about any errors, which are all my
fault and not the fault of the speaker. 🙂