SCAT: Structure-Conditioned Antibody Transformer for CDR Design

Antibody engineering is a cornerstone of modern therapeutic development, yet the rational design of antigen-specific antibody CDR sequences remains computationally demanding. We present SCAT (Structure-Conditioned Antibody Transformer), a causal masked-infilling framework for antigen-conditioned antibody CDR sequence generation. SCAT combines a large antibody repertoire prior with an SE(3)-invariant geometry encoder for antigen structure and multi-layer antigen cross-attention in the decoder. We test whether structural conditioning measurably influences sequence generation using controlled interventions, including antigen substitution and systematic degradation of antigen geometry. Both experiments identify antigen-dependent CDR-H3 positions whose residue preferences change with antigen input, while conserved loop-terminal positions remain comparatively stable. On CHIMERA-BENCH, SCAT achieves an AAR-H3 of $0.449$, numerically above values reported for five structure-conditioned CDR-design baselines on the same benchmark; differences in training corpora, inference inputs, and task formulation preclude a controlled head-to-head comparison. For two held-out reference complexes, HER2 and TROP2, we further subject generated candidates to a post hoc computational screen using Boltz-2 complex prediction and FoldX interaction-energy scoring. These scores are not fed back into generation and are treated as prioritization proxies rather than evidence of binding or specificity.