AF3Design: Selectivity-Aware Nanobody Design with AlphaFold3

Abstract

Immunotherapy, which harnesses a patient's own immune system to fight disease, has emerged as a promising approach to cancer treatment. However a critical limitation of this therapeutic is low selectivity, which results in possible damage to healthy tissue and other adverse side effects. We present AF3Design, the first computational framework for nanobody design that integrates binding selectivity directly into the optimization process rather than relying on post-hoc filtering. Our approach leverages AlphaFold3's structure prediction capabilities through a gradient-free genetic algorithm, optimizing a composite fitness function with explicit off-target penalties. These terms include contact-based rewards for discriminating residues and binding energy predictions. We validate our method on the challenging task of designing nanobodies against peptide-MHC bearing healthy or cancer-associated peptides that differ by as little as a single residue. AF3Design significantly improves peptide-facing interface confidence over baselines and enriches binders that preferentially contact the discriminating peptide residue(s), with further binding energy gains after staged activation. Our findings demonstrate that selectivity-aware optimization is a tractable and effective strategy for engineering nanobodies capable of distinguishing near-identical peptide antigens.