Bacteria exhibiting antimicrobial resistance (AMR) is a problem that has grown to become a significant global public health challenge. Antibiotic resistance genes (ARGs), determinants of AMR, mostly emerge from non-clinical settings. Identifying novel ARGs from the human microbiome which confer resistance to clinical antibiotics is a crucial component of addressing AMR. We sought to leverage the increased accuracy of computational protein structure prediction by training a one-class support vector machine on the pairwise primary protein structure and tertiary protein structure distributions of conserved structural regions of various ARG classes. We applied the computational method to seven human microbiome project reference strains and functionally confirmed eight out of nine tested novel ARG predictions, belonging to ARG classes: DFR, class B and C β-lactamases and penicillin binding proteins. We also applied the method directly to protein structures within the AlphaFold database and functionally confirmed a novel metallo-β-lactamase with low homology to existing ARGs. In total, 70% of tested genes confer resistance to clinical antibiotics at CLSI resistant breakpoints, or are emerging from the pre-resistome and may become clinically relevant in the future. This study represents a precise and computationally efficient method of identifying previously uncharacterized ARGs from DNA databases.