Abstract: Predicting protein binding affinities across large combinatorial mutation spaces remains a critical challenge in molecular biology, particularly for understanding viral evolution and antibody interactions. While combinatorial mutagenesis experiments provide valuable data for training predictive models, they are typically limited due to experimental constraints. This creates a significant gap in our ability to predict the effects of more extensive mutation combinations, such as those observed in emerging SARS-CoV-2 variants. We present ProxiClust, which strategically combines smaller combinatorial mutagenesis experiments to enable accurate predictions across larger combinatorial spaces. Our approach leverages the spatial proximity of amino acid residues to identify potential epistatic interactions, using these relationships to optimize the design of manageable-sized combinatorial experiments. By combining just two small combinatorial datasets, we achieve accurate binding affinity predictions across substantially larger mutation spaces ($R^2\approx0.8$), with performance strongly correlating with the capture of high-order epistatic effects. We validate our method in five different protein-protein interaction data sets, including binding of SARS-CoV-2 receptor binding domain (RBD) to various antibodies and cellular receptors, as well as influenza RBD-antibody interactions. This work provides a practical framework for extending the predictive power of combinatorial mutagenesis beyond current experimental constraints, offering applications in viral surveillance and antibody engineering.