Rare variants are of increasing focus in genetic association studies. Standard aggregation analyses improve association signals but lack variant-specific details that could be crucial to the follow-up studies. Localizing rare risk variants is challenging because the information content of each rare variant is too little to separate signals from the noises. In this work, we propose a structure-guided test that incorporates protein tertiary structure to guide local variant collapsing and assess single RV association. Protein tertiary structures contain constructive information on how variants interact and function together, and enable information borrowing from variants that are close in structural space but are apart in sequence location. The proposed test is built on a structural kernel that incorporates protein structure, adaptively determines the proper amount of information borrowing from neighboring variants, and uses resampling methods to obtain variant-level significance. We demonstrate the utility of the structure-supervised test of single RVs using simulations and real data applications.