We added genes one at a time by picking the most selective gene, in combination with the previously selected genes. When more than one gene was equally selective, we picked the one that was most highly enriched. We defined selectivity as the reciprocal of the number of clusters that would be assigned auto-annotation given the current gene set and the trinarization scores. That is, an annotation that would apply to k clusters would have selectivity 1/k. Adding more genes rapidly drove selectivity toward 1.
