Proteins are biomolecules that are of great importance in science, biotechnology and medicine. Their function relies heavily on their three-dimensional shape, which in turn follows from their amino acid sequence. Therefore, there is great interest in modelling the three-dimensional structure of proteins in silico given their sequence. We discuss the formulation of a tractable probabilistic model of protein structure that features atomic detail and can be used for protein structure prediction. The model unites dynamic Bayesian networks and directional statistics to cover the short-range features of proteins. Long-range features are added by making use of probability kinematics - a little known variant of Bayesian belief updating first proposed by the probability theorist Richard Jeffrey in the 1950s. The method we describe can be generalized to formulate tractable probabilistic models that involve high dimensionality and need to cover multiple scales