• Encode antibody and antigen with ESM2-nv (ESM2 but on NVIDIA), concatenate embeddings and feed into a lightweight transformer (4 attention heads, 7 layers) to predict binding affinity
• Pretrain AlphaBind on 7.5M antibody-antigen affinity measurements from AlphaSeq (yeast display), then finetune for each parent antibody using ~30K variants
• Use ESM2-nv masked prediction to guide antibody mutations, filter candidates with AlphaBind affinity predictor