Figure 1

Bias correction improves concordance of replicates in a large breast cancer microarray study. Hierarchical clustering was performed, and the number of replicate samples correctly paired (joined at the lowest node) was used as a measure of concordance. Colors underneath the dendrograms indicate replicate samples in which each pair was taken from a single tumor specimen, and black squares indicate samples taken from normal breast. (a) The unfiltered, uncorrected data results in only four of nine replicate pairs together. (b) Filtering out all but the top 100 genes with highest variance results in all 9 pairs clustered together. (c) Bias correction as described in this paper also results in all nine replicate pairs clustered together, but in this case all genes are retained. (d) The concordance achieved by variance filtering is sensitive to the number of genes retained. Generally, filtering out low-variance probe sets with increasing stringency leads to increased concordance between replicate sample pairs.