Advancing AI for biology: Teaching models to design and characterize antibodies
Source
www.amazon.science
Date
Key takeaways · AI-distilled
MochiBind, Amazon's antibody-binding predictor, reframes the task from predicting absolute binding affinity to ranking which of two candidates binds more tightly — a pairwise approach the team found more robust across assays than absolute-value predictors.
Tested on four held-out antigens (TIGIT, PD-1, HER2, SARS-CoV-1 RBD) with strict cross-antigen validation, MochiBind beat every structure baseline on pairwise accuracy by ~10% average, and scored 200,000 antibody pairs in ~13s on a CPU — 100x faster than rivals.
CA-MAP, a second predictor, corrects for lab-to-lab batch effects via in-context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → examples from the same lab rather than retraining: under a simulated batch effect, standard fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition →'s correlation with ground truth dropped from 0.99 to 0.58, while CA-MAP held at 0.99.
CA-MAP is dramatically cheaper than the baseline it beats: about 182,000 trainable parameters versus TxGemma's 40 million, and roughly 200 times faster per prompt at inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →.
In a campaign against a novel pediatric-cancer target with no prior antibody data, Amazon's hotspot-AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → and three generative models produced 288,000 candidates; after yeast-display screening and two sorting rounds, 46 of 116 survivors were confirmed strong, specific binders.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
