This graph shows that links that show high intra-dataset confirmation (see Supplementary figure B) also tend to show high inter-dataset confirmation. The curve for "all links" here is essentially the same data shown in Figure 2A in the paper, plotted as a cumulative probability.
"Worst" were those that were tested at least 10 times in a data set, but selected only once (e.g., a confirmation rate less than 0.1). The "best" links were those that were tested at least 4 times and yielded confirmation rates of 1.0. This determination is made on a per-data set basis, so a link might have a high intradataset confirmation rate in one data set but a low rate in another. In this situation, we use the best confirmation rate to represent the link. The graph thus shows that links that tend to have low intra-dataset confirmations also tend not to be confirmed by other data sets. One practial effect this has is that even though the "worst" data sets are numerous, they have a relatively small effect on the links we detect as being confirmed between data sets.