Answers are grounded in this paper and cite their sources.
Structure-based machine learning is widely expected to advance antibody engineering, yet it is unclear how far affinity predictions transfer to new targets, and why. We introduce a noise-corrected, censoring-aware estimator of the cross-target transfer ceiling, ρ max = √(T/(T+I)), which separates a transferable antibody-intrinsic component (T) from a non-transferable target-specific interaction (I). On deterministic synthetic ground truth the estimator recovers the noise-free ceiling to ± 0.0003, whereas the naive sum-of-squares estimator is biased downward by measurement noise. Applying it from synthetic to real systems, the transferable budget is small, and on real neutralization panels it is not an artefact: the target-specific interaction survives noise and censoring correction. It is strongly regime-dependent, with a two-tier structure: predicting a novel antibody from sequence alone (the design setting) reaches a lower ceiling (ρ ≈ 0.44 on real breadth data) than ranking a characterized one (ρ ≈ 0.59), even with a strong protein language model. Synthetic oracles also overstate transferability: an oracle scored on real predicted structures carries about half the budget of a matched synthetic lattice (0.45 vs. 0.81), and real neutralization and escape panels sit well below the synthetic ceiling (0.32-0.68). Under leakage-free antigen-cluster splits, cross-target affinity ranking across every method family collapses to this budget, while within a sampled target a structure-based graph network ranks affinity above off-the-shelf FoldX and GearBind (within-complex Spearman 0.55), and structural surface descriptors predict a surface-driven developability property (hydrophobic aggregation) where sequence-dominated properties (humanness, charge-driven self-association) do not benefit. A controlled ablation attributes the residual within-target signal to the antibody alone: dropping the antigen from the structural model leaves it essentially unchanged (interface Δ ρ between -0.01 and +0.02 across four panels), so for this model the paratope-epitope interface adds no ranking signal over the antibody alone. Models climb toward the budget within a target — via a measure-then-rank operating mode that recovers top binders 6-8 times more efficiently than random once a target has been sampled — but none crosses it.
No takes yet. Share an insight, caveat, or question.
Martins et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: