had little effect on the binding signal compared to sites with larger PWM score changes (Figure 5b), suggesting that the PWM model adequately describes the functional constraints of CTCF binding sites. At the same time, we found that CTCF binding signals could be maintained even in those cases where mutations resulted in significant changes of PWM score, particularly at evolutionarily conserved sites (Figure 5c). A linear interaction model confirmed that the effect of motif mutations on CTCF binding was significantly reduced with increasing conservation (Figure 5d; interaction term P = 2.9e-2). These effects were not due to the presence of additional CTCF motifs (as 96% of bound regions contained only a single motif), while differences between more and less conserved sites could not be explained away by differences in the PWM scores of their major alleles (not shown). A CTCF dataset from three additional individuals generated by a different laboratory [44] yielded consistent conclusions (Figure S6B-D in Additional file 1), suggesting that our observations were not due to overfitting.
