It has become clear that there is no simple ‘code’ for protein–DNA recognition and that selecting an optimal binding sequence along the DNA double helix corresponds to more than simply forming a set of specific hydrogen bonds or steric interactions. However, it has been difficult to characterize the so-called indirect components of recognition. While DNA deformation certainly underlies indirect recognition, it is not easy to determine how local fine structure and deformability depend on base sequence or exactly what percentage of recognition should be attributed to such factors. Molecular modelling can help to develop these ideas into a quantitative model, provided the calculations can be carried out fast enough to enable a comprehensive survey of base-sequence effects. I present here some recent results from our group and their consequences for improving our understanding of protein–DNA binding, and their potential for predicting, and eventually modulating, protein–DNA binding.