Comparative Analysis of Collapsing Methods

For each of our examples, we ran a subset of the following six collapsing methods and studied how often each method leads to the most reproducible results. First, we choose the row with the highest mean expression (1.max) by setting method = "MaxMean" and connectivityBasedCollapsing = FALSE. Second, we choose the row with the highest between-column variability (2.var) by setting method = "maxRowVariance" and connectivityBasedCollapsing = FALSE. Third, we choose the row with the highest connectivity in cases with three or more rows per group or highest mean expression in cases with two rows per group (3.kMax) by setting method = "MaxMean" and connectivityBasedCollapsing = TRUE. Fourth, we choose the row with the highest connectivity in cases with three or more rows per group or maximum variability in cases with two rows per group (4.kVar) by setting method = "maxRowVariance" and connectivityBasedCollapsing = TRUE. Fifth, as a sort of control, we compare our results to a standard method of centroid determination by measuring the module eigengene (first principal component) for all rows in a given group across all groups (5.ME) by setting method = "ME". Finally, in our assessment of blood cell type, we also use the average of all marker genes (6.Avg) by setting method = "average".