Abstract
The prediction of antibody binding residues of an antigen is essential for understanding the immune response mechanisms and advancing antibody therapeutics. By definition, each epitope is the binding site of a specific antibody; however, many prediction methods are antibody-agnostic, and thus need only the structure of the antigen. Antibody-specific methods also require either the structure or models of the antibody, and are generally based on docking or co-folding algorithms. Machine learning methods have been improved substantially during the last few years, resulting in new approaches to epitope prediction. We evaluate some popular methods and show that combining AlphaFold 3 with the epitope prediction program AbEMap yields substantially better results than any of the other methods tested.