The promoter regions from the reference set were analyzed using the same procedures as were applied above (Figure 2d-f). In spite of the likelihood that the new reference collection will have greater noise than the small set collected by detailed literature analysis, the performance results are comparable between sets. The sensitivity is slightly lower for the large collection (Figure 2e,f), which in addition to the potential difference in annotation standards could be attributable to the TFs associated with the sites. The average information content of the models for TFs linked to sites in the reference collection is lower than that for the factors associated with the small test set (median information content: 9.7, as compared to 15.3 bits in the first test set). Selectivity performance is virtually identical to the test (Figure 2d,f).