Generative Language Modeling for Antibody CDR Grafting and Alignment

Antibodies recognise their targets through hypervariable complementarity-determining regions (CDRs), which are interleaved with conserved frameworks in sequence space, making de novo CDR design an infilling problem. Autoregressive models generate residues left to right, precluding full framework context during CDR generation and conflating framework and CDR likelihoods, with no natural prompt–response interface for feedback to steer generation. We introduce GenCDR, a family of framework-first antibody language models that use all frameworks as a conditioning prompt and jointly generate CDRs as a variable-length response. The family comprises IgGenCDR, p-IgGenCDR, and NanoGenCDR, trained on unpaired, paired, and nanobody chains, respectively. GenCDR achieves the highest CDR recovery among autoregressive models and generates natural, diverse, human-like CDRs whose likelihoods correlate with fitness and developability assays. The prompt–response boundary also enables principled alignment: reward signals for binding affinity, expression, or developability can be composed to steer CDR generation. In an iterative loop coupling the generator to a co-folding oracle and developability screening, we use reinforcement learning to align NanoGenCDR generation towards a specific antigen. After alignment, NanoGenCDR achieves in silico interface metrics competitive with a structure-conditioned diffusion pipeline using roughly half the sampling budget and producing more natural and developable designs. The antigen is never explicitly encoded, demonstrating that co-folding feedback can effectively steer GenCDR towards specific, antigen-binding CDRs.