Computational protein design has made substantial progress over recent years generating new enzymes, binders, and inhibitors not seen in nature. Among still unsolved challenges are design of new backbones for function and design of membrane-protein interactions. We have developed new algorithms for design of backbones from modular pieces of natural proteins and applied this strategy to design new antibody binders of insulin. The antibodies, which are distant from any natural germ-line antibody by at least 50 mutations, bind at mid-nanomolar dissociation constants and mutations introduced through experimental affinity maturation appear to rigidify the bound conformation. Using a similar approach we also designed the first antibody binder which can be controlled allosterically through binding of a small-molecule effector. The resulting antibody binds its target with two orders of magnitude higher affinity in the presence of the effector than in its absence. In our analysis the major hurdle towards consistent design of membrane proteins remains the uncertainty regarding the energetics of membrane-protein insertion and association in the membrane. We have extended the popular TOXCAT assay for probing membrane-protein homodimerisation to a high-throughput assay able to monitor the effects of every point substitution on membrane-protein insertion and dimerisation in the bacterial inner membrane in a single experiment. The result of applying this method to model membrane segments is a per-position map of the mutational landscape of a membrane protein, from which we can extract thermodynamic quantities for membrane insertion and association. This method is the first, to our knowledge, to experimentally quantify the effects of the positive-inside rule and suggests ways of stabilising membrane proteins directly from sequence. Together with ab initio modelling we are able to use the resulting homodimerisation energy maps to generate atomic-accuracy models of membrane-protein homodimers, including of the ErbB2 oncogene.