Computational methods that predict the structure and specificity of protein-protein interactions can yield deep insight into the structural biology of many biochemical pathways. Through high-resolution structures of protein-interactions we can identify the structural mechanisms of diseases, engineer proteins towards specific functions, and design drugs that disrupt pathogenesis. Challenges in accurately modeling protein interactions include efficiently sampling the conformational space available for two proteins to interact, and adequately approximating the free energy of the conformational landscape to correctly predict the structure and specificity of the protein interaction. In this thesis I detail my work on developing new methods in flexible protein and peptide docking and applying it in a number of areas to both predict the atomic-scale structure of protein interactions as well as model their specificity. First I introduce our early efforts in predictive protein docking and demonstrate the limitations of the prevailing rigid-body approach to docking while outlining the challenges in modeling binding-induced conformational changes in docking. Second, I introduce a new approach to modeling flexibility in protein docking called ensemble docking, inspired by the conformer selection model for protein binding. I apply this method towards modest conformational changes in crystal structures and then extend it towards docking with NMR structures and homology models. Third, I use flexible docking to model interaction specificity of the enzyme HIV-1 protease. I develop a flexible peptide docking algorithm that predicts the structure of the enzyme-substrate complex and calculates the energetics of the enzyme-substrate interaction and use statistical analysis and energy decomposition to identify enzyme residues that are most important for substrate specificity. Fourth, I look at our efforts towards the future of modeling protein-protein interactions, first through the newly updated RosettaDock v3 and through PyRosetta, a Python script-based implementation of Rosetta that is intended to make molecular modeling more accessible to the biomedical community. In summary, I have developed novel flexible protein docking methods that have had success in modeling the structure and specificity of protein interactions in a wide-range of applications, from predictive protein docking, antibody-antigen homology modeling and docking, and the identification of specificity determinants in enzyme-substrate interactions.