These limitations, however, are universal for the problem of modeling functional constraints based on sequence alone. The predictive power of PWMs is probably comparable with our ability to predict the impact of mutations on RNA and protein structure. The rapidly increasing bulk of genotyping data will increase the statistical power of these analyses, but only experimental validation of the effects of TFBS mutations can give a definitive answer. This is why direct analyses of TF binding across individuals hold much promise. Using multi-individual CTCF binding maps [16,44], it was reassuring to confirm that the loss of CTCF binding associated with a TFBS mutation is generally proportionate to its impact on motif PWM match. But perhaps more importantly, using these data has allowed us to observe that this relationship does not always hold, suggesting that variation at many sites, and in particular the most evolutionarily conserved ones, can be efficiently buffered at the binding level. We do not know the exact nature of these buffering mechanisms, and whether their prevalence at highly conserved sites is evolutionarily driven or is merely a
