Advancing AI for biology: Teaching models to design and characterize antibodies

Amazon Bio Discovery, described as an AI-powered application giving scientists access to biological AI models and integrated lab services, has published three papers on AI-assisted antibody design and characterization. Two are peer-reviewed journal papers on prediction; the third applies prediction within an end-to-end design process and reports experimentally validated antibody hits against a novel cancer target.

The first paper, "A systematic evaluation framework for universal antibody-antigen binding affinity prediction and candidate recommendation," published in iScience, introduces MochiBind, a sequence-based predictor. Rather than predicting absolute binding affinity, MochiBind predicts which of two antibodies against the same antigen binds more tightly, embedding antibody-antigen complexes with the ESM-2 protein language model and aggregating pairwise comparisons into a global ranking using TrueSkill. No structural input is required. Evaluated on the AlphaBind dataset across four antigen systems (TIGIT, PD-1, HER2, and the SARS-CoV-1 RBD), MochiBind achieved higher pairwise accuracy than every structure-based baseline on all four held-out antigens, outperforming the closest competitor by almost 10% on average. It scored 200,000 antibody pairs in roughly 13 seconds on a CPU, described as a more than 100-fold inference speedup.

The second paper, "Context-aware multi-property antibody predictor: A novel framework integrating text and protein language models," in npj Systems Biology and Applications, presents CA-MAP, which addresses batch effects during inference by taking a prompt of example antibodies with measured properties. Its AB-context-aware training strategy applies a hidden random transformation to context properties and expected answers. In hydrophobicity prediction on a fine-tuned TxGemma model, the context-aware approach retained a 0.99 Spearman correlation under a simulated additive batch effect in the 0–0.3 range, where standard fine-tuning fell to 0.58. CA-MAP uses roughly 182,000 trainable parameters versus TxGemma's 40 million and is about 200 times as fast per prompt. Trained on a synthetic dataset of 876,898 antibody-heavy chains covering six developability properties, it reached a Spearman correlation above 0.8 on several properties.

The third paper, "Agent-guided de novo design of nanobody binders against a novel cancer target," presented as a Spotlight at the ICML 2026 Workshop on Generative and Agentic AI for Biology, where it received the Best Paper Runner-Up Award, targeted a cell surface antigen for desmoplastic small round-cell tumors, a rare and aggressive pediatric cancer, identified by collaborators at the Dr. Nai-Kong V. Cheung's Lab at Memorial Sloan Kettering Cancer Center. The target has no experimental structure and no public antibody information. A hotspot recommendation agent proposed eight hotspot regions; three generative models (RFantibody, IgGM, and mBER) each produced 96,000 designs, and a selection agent prioritized 100,000 candidates for yeast-display screening. None of the 116 surviving candidates bound an unrelated control protein, and 46 were identified as strong binders.