Antibody-antigen interactions are critical to our immune response, and understanding the structure-based biophysical determinants for their binding specificity and affinity is of fundamental importance. We present a computational structure-based cross-docking study to test the identification of native antibody-antigen interaction pairs among cognate and non-cognate complexes. We picked a dataset of 17 antibody-antigen complexes of which 11 have both bound and unbound structures available, and we generated a representative ensemble of cognate and non-cognate complexes. Using the Rosetta interface score as a classifier, the cognate pair was the top-ranked model in 80% (14/17) of the antigen targets using bound monomer structures in docking, 35% (6/17) when using unbound, and 12% (2/17) when using the homology-modeled backbones to generate the complexes. Increasing rigid-body diversity of the models using RosettaDock's local dock routine lowers the discrimination accuracy with the cognate antibody-antigen pair ranking in bound and unbound models but recovers additional top-ranked cognate complexes when using homology models. The study is the first structure-based cross-docking attempt aimed at distinguishing antibody-antigen binders from non-binders and demonstrates the challenges to address for the methods to be widely applicable to supplement high-throughput experimental antibody sequencing workflows.