Protein-protein interactions underlie virtually all biological processes, from intracellular signaling to immune recognition. Although recent advances in deep learning have substantially improved structure prediction and canonical protein interaction modeling, many biologically important interactions lack strong sequence homology or explicit coevolutionary signals, limiting the applicability of existing approaches. This thesis develops computational frameworks to model molecular recognition beyond classical approaches and to translate these insights into immunoengineering. In Chapter 2, we introduce MaSIF-mimicry, an interface-level similarity framework for discovering latent protein interactions using interface mimicry, independent of sequence or structural homology. Our analysis of the human interactome reveals that the protein interface space is highly degenerate, with a limited repertoire of reusable interaction patterns that enable a single protein to engage multiple partners. Leveraging this property, MaSIF-mimicry identifies cross-species host-pathogen interactions and uncovers novel neosubstrates recruited to the E3 ligase CRBN in the presence of molecular glue degraders. These results demonstrate that interaction compatibility can be inferred from shared interface geometry and chemistry even when classical coevolutionary evidence is absent. In Chapter 3, we extend this perspective to antibody-antigen recognition. Using a curated dataset of HIV Env EMPEM models, we train a machine learning model to learn epitope surface features that govern antibody targeting. We show that immunogenicity is heavily shaped by an epitope's intrinsic features, including its geometry and amino acid composition. These determinants explain immunodominance hierarchies and enable the rational design of immunofocusing antigens that redirect antibody responses toward conserved, subdominant epitopes. In Chapter 4, we integrate insights from previous chapters into an evolutionary sequence modeling framework for antigen design. We use protein language models to analyze the evolutionary history of the influenza hemagglutinin and learn the rules governing antigenic evolution. Our framework implicitly encodes constraints on residue composition, structural context, and antibody engagement potential, as well as mutational pathways associated with immune escape. Modeling trajectories within this latent space enables the generation of novel antigen sequences that incorporate targeted antigenic features, providing a principled strategy for evolution-aware immunogen design. Together, this thesis establishes a unified framework for analyzing, discovering, and designing molecular recognition. By integrating structural and evolutionary representations of proteins, this work advances approaches for PPI discovery, antibody-antigen interaction modeling, and vaccine design.