(A) LOOCV performance of each of nine different methods on 31 data sets in benchmark. Color accents represent CRM-level sensitivity on a scale of 0 to 1, and cases with empirical p-value ≤ 0.05 are marked by asterisks. The top row shows the number of data sets amenable to supervised prediction by each method. One-on-one comparison of methods: For each pair of methods M1 (row) and M2 (column), (B) the “wins” of M1 versus M2 (i.e., the number of data sets on which CRM-level sensitivity of M1 was greater than that of M2 by at least 10% of data set size). (C) the difference between the wins of M1 versus M2 and the wins of M2 versus M1 in CRM-level sensitivity (D) Fifteen data sets on which at least one method succeeds in all four instantiations. Color indicates the number of instantiations (out of four) on which the performance was significant (p ≤ 0.05): white=4, yellow=3, orange=2, brown=1, black=0. (E) Comparison of single species and multi-species versions of HexMCD. For each each LOOCV instantiation, the average CRM-level sensitivity over all data sets (y-axis) and number of amenable data sets (number above each bar) are compared between the two methods.

Each predicted CRM was used to drive a GFP reporter gene in transgenic embryos. Expression was visualized using antibodies to GFP and compared to the endogenous expression of the putative associated gene as determined by whole mount in situ hybridization to mRNA (in the case of odd, Odd antiserum). All embryos are shown with anterior to the left. (A) The edl CRM drives expression in the anterior of blastoderm stage embryos consistent with endogenous edl expression (B). At embryonic stage 10, reporter gene expression (C) mimics edl gene expression (D) at the ventral midline. By stage 11, reporter gene expression is no longer seen at the ventral midline despite continued endogenous gene expression at this site (arrowheads in E and F); however, reporter gene expression is maintained in the developing brain (arrows in E and F). The srp CRM reporter gene (G) is expressed similarly to srp mRNA (H) in the blastoderm as well as in stage 8 embryos in the developing posterior endoderm (I, J, arrowheads) and anterior endoderm (I, J, arrows). The odd CRM recapitulates odd gene expression in mesodermal progenitors (K, magenta in L) but not in the ecotderm (L; anti-Odd is in green and is nuclear, GFP expression driven by the Odd CRM reporter is in magenta and primarily cytoplasmic). Inset in L shows a close up of the mesodermal cluster marked by the arrowhead. (M-P) The SoxN CRM drives reporter gene expression in a subset of the endogenous mRNA pattern in the central nervous system. At stage 9, reporter gene expression is visible in the brain (M, arrow) but not the remainder of the central nervous system (M, arrowhead; compare with N). By stage 14, additional nervous system expression is apparent (O), consistent with endogenous SoxN (P). (Q-T) cas CRM reporter gene expression faithfully reproduces cas expression in the late embryonic ventral nerve cord and brain. Nerve cord expression of both the reporter gene and the endogenous mRNA is most prominent in a set of lateral cells in the thoracic segments (arrows in Q, R); additional reporter gene expression in midline cells (Q, arrowhead) is likely perdurance from a slightly earlier stage of expression (data not shown). Brain expression of the reporter gene, like that of cas mRNA, is strongest in a group of medial cells (S, T, arrows), with weaker expression laterally (S, T, arrowheads).

LacZ reporter constructs were microinjected into mouse embryos and assayed by wholemount staining of midgestation F0 embryos. Shown to the left is a transgenic embryo with the LYL1 promoter construct showing transgene expression in heart (read arrowhead), fetal liver (purple arrowhead) and vessels (green arrowhead). The middle and right hand panels show representative transgenic embryos with reporter constructs for the predicted CRMs in exon 3 of C1ORF164 and exon 6 of EBF3 respectively. Arrowheads indicate staining in heart, fetal liver and vessels.

(A) LOOCV performance of each of seven different motif-blind methods on 8 data sets in benchmark. Format is identical to that of Fig. 1A. (B) Consistency of a method’s accuracy on each data set, over ten instantiations of LOOCV. Colors and numbers indicate the number of instantiations on which the performance was significant (at p ≤ 0.05). (C) The set of nearest genes for the top 1000 CRM predictions for early blood and cardiovascular development in human/mouse was intersected with a set of 7035 genes differentially expressed in blood stem cells in mouse (Miranda-Saavedra et al., 2009). The intersection size (“Predictions”) was significantly above chance expectation (“Z-score”; mean 204, standard deviation 13, estimated by simulations). (D) Predicted CRMs were scanned for significant occurrence of Ets-related and GATA motifs from TRANSFAC, counts of motif-containing predictions are in Column 3, and significance estimates of these counts, based on random sampling of CRMs, are in Column 4.