Developability-Guided Antibody Design Using FlashTAP

Antibody design usually treats developability as a filter rather than an optimisation objective. This means designs which may otherwise have ideal binding properties might still be rejected due to developability issues. We integrate FlashTAP, a differentiable predictor for four classical developability metrics, at three inference-time intervention points: post-hoc repair, structure-guided inverse folding and de novo backbone design. On a fixed 34-antibody panel carrying 71 native classical-TAP risk flags, post-hoc repair cleared up to 85% of them with the whole Fv mutable and up to 38\% with only the framework mutable. Steering a diffusion-based inverse-folding model with developability gradients doubled the flags it cleared, from 15% to 32%, whilst preserving fold consistency and naturalness. In de novo design, guidance cut the mean flag count of hallucinated backbones by 58% per design. Overall, we show that developability-driven guidance can be used to directly optimise antibody sequences at various stages of the design pipeline.