Table 2.

The Performance of the Various Sampling Modes in the Prediction of the Sequence-Specific Myf, Mef2, and SRF Sites

	Total no. of reported Mef2, Myf, and SRF sites	Total no. of predicted Mef2, Myf, and SRF sites	No. matching reported Myf, Mef2, and SRF sites	% of predicted sites overlapping reported sites	No. of sites predicted (includes predicted SP1 model and unknown model)
Motif (no mouse)a	48	123**	13	27.1	369
Motifb	96	72	48	50	132
Module (no mouse)c	48	9*	0	0	109
Clustered sitesd	96	112	52	54.2	222
Modulee	96	104	66	68.75	176

To examine the importance of the alignment of homologous sequences in finding regulatory modules and the role of clustering and neighboring interaction, we compared five versions of the algorithm: ^aRow 1: the motif sampler similar to the one used by Wasserman et al. (2000) (no restrictions on the clustering of sites and no neighboring effects) applied to the human sequences only; ^bRow 2: the motif sampler modified to sample simultaneously from aligned sequence pairs; ^cRow 3: the module sampler applied to the human sequences only; ^dRow 4: the module sampler applied to aligned human mouse sequence pairs but with the neighboring interaction component inactivated (thus, yielding a model that enforces clustering but has no neighboring effects); ^eRow 5: the full module sampler, including both clustering and neighboring applied to aligned human mouse sequence pairs. In all cases, we searched for five different models simultaneously. The ^* indicates that the algorithm did not predict a Mef2 or SRF-like model, but only a Myf-like model. ^** indicates that two different weak Myf-like models were predicted.