Perfect conservation (i.e., 100% confidence in ancestral position identity), would result in a large reduction in dataset size due to the "false negative" exclusion of many positions due to the presence of these occasional derived states.
At each step in the reduction, the dataset was reduced by half; by the 15th iteration the data was reduced to a total of 15 CpG sites.
In some cases, the overall reduction in dataset size per locus was considerable (e.g., 21.9% [ 50], 13.6% [ 51], up to 8.5% [ 52]).
Despite the important reduction in dataset size, focusing on the most informative genes (2545 genes with at least four polymorphic sites) allowed us to compare the gene FI distribution in 105 different functional classes, each containing at least 40 different genes.
This splitting does not cause linear slowdown; there is simply more overhead of launching, so datasets with large numbers of instances see less of a performance reduction than datasets with few instances.
Among the statistical techniques, both principal component analysis (PCA) and cluster analysis (CA) are useful methods to discover common patterns in data distribution, leading to initial dimension reduction of datasets and helping its interpretation (Franco-Uría et al. 2009).
Several previous studies have considered the importance of similarity reduction in datasets of MHC-II peptides.
Principal components analysis was performed for data reduction, simplifying datasets to three dimensions for plotting purposes.
The first pre-processing stage aims at noise reduction within each dataset, to prepare it for cross-platform normalisation.
One major cause of the reduction of the dataset is that we were only able to work with genes that are duplicated in A. oryzae.
