Figure 2.

Initially confident specificities tend not to change. For two independent datasets of single-domain C2H2-ZF specificities, we apply the QP formulation to each dataset separately for different values of α (x-axis). (Top) For each α, for all proteins shared between the two datasets, we compare their jointly inferred specificities in each of the two datasets and compute the fraction (y-axis) of initially disagreeing columns that now agree (green) and the fraction of initially agreeing columns that now disagree (red). For all α, agreement gain is substantially larger than agreement loss. (Bottom) We plot the ratio of these two values (green over red, or the enrichment; y-axis), observing 4-to-8-fold enrichment for agreement gain, peaking near α = 0.4. We note that that large enrichments at high α (≥0.90) are an artifact of small sample sizes (i.e. most columns’ agreement statuses have not yet changed from their initial status; see top).