Date:
Type:
Publication:
Author(s):
Computational antibody design offers an appealing alternative to experimental affinity maturation. Mutations, insertions, and deletions frequently induce substantial conformational rearrangements in antibodies, whereas many Free Energy Perturbation (FEP) applications are restricted to relatively small changes between structurally similar molecules. Here, we introduce Tango for biologics, an implicit-solvent absolute binding free energy (ABFE) framework that enables whole-protein alchemical transformations between antibody variants without requiring shared scaffolds or structural similarity. Tango was evaluated on three benchmark antibody–antigen series, using both crystallographic and co-folded structures. On crystal inputs, Tango recovered experimental rank ordering of binding free energies with Spearman correlations (ρ) ranging from 0.39 in charged, polar environments to 0.75 in hydrophobic interfaces. A predictive signal was obtained from high-quality co-folded structures, yielding ρ ranging from 0.43 to 0.58. Further, Tango achieved robust lead prioritization across all benchmark systems, with normalized discounted cumulative gains of 0.60–0.83. Co-folding failures were primarily driven by inaccurate antigen placement rather than monomer folding, even for complexes represented in co-folding training sets. Refinements to co-folding methods partially rescued difficult benchmark systems, enabling the generation of simulation-ready complexes. Together, these results define the conditions under which AI-derived structures are sufficient for rigorous affinity prediction, when only lead prioritization remains reliable, and when current co-folding approaches fail prospectively.
