What does AlphaFold3 learn about antibody and nanobody docking, and what remains unsolved?

Antibody therapeutic development is a major focus in healthcare. To accelerate drug development, significant efforts have been directed toward the in silico design and screening of antibodies for which high modeling accuracy is necessary. To probe AlphaFold3's (AF3) capabilities and limitations, we tested AF3's ability to capture the fine details and interplay between antibody structure prediction and antigen docking accuracy. With one seed, AF3 achieves a 10.2% and 13.3% high-accuracy docking success rate for antibodies and nanobodies, respectively. AF3-like models Boltz-1 and Chai-1 achieve 4.08% and 0% high-accuracy rates for antibodies, and 5% and 3.33% for nanobodies, respectively. With twenty seeds, AF3 achieves a median unbound CDR H3 RMSD accuracy of 2.9 Å … and 2.2 Å … for antibodies and nanobodies, respectively. Both AF3-like models Boltz-1 and Chai-1 improve further on antibodies (2.08 Å … and 2.71 Å …, respectively), but do poorly on nanobodies (3.78 Å … , 3.63 Å …). CDR H3 accuracy boosts AF3 complex prediction accuracy, with antigen context improving CDR H3 accuracy, particularly for loops longer than 15 residues. Combining ipTM-HA and I-pLDDT with $\Delta G_B$ improves discriminative power for correctly docked antibody and nanobody complexes. However, AF3's 65% failure rate for antibody and nanobody docking (with single seed sampling) demonstrates a need to further improve antibody modeling tools.