Amazon's AI Just Designed 46 Cancer-Fighting Antibodies From Scratch

Amazon's AI-for-biology team says it has taught computers to do one of the slowest jobs in medicine: designing antibody drugs. Antibodies are the body's targeted missiles — proteins that latch onto one specific invader — and more than 100 have become FDA-approved medicines for cancer, arthritis and infections. Today, going from a promising target to a candidate worth testing in the lab takes roughly six to twelve months and a lot of money. Amazon Bio Discovery published three papers describing AI that speeds up each step.

The first tool, MochiBind, doesn't try to guess exactly how tightly a drug will stick. It simply ranks thousands of candidates and picks the best ones — which is what scientists actually need. Amazon says it's 10% more accurate than competing tools and 100 times faster, scoring 200,000 candidates in 13 seconds on an ordinary computer chip. The second tool, CA-MAP, predicts whether a candidate can be manufactured at scale, even when data from different labs don't line up — a common reason promising drugs quietly die.

The most striking part: a third system using AI "agents" (software that can take steps on its own) chose where to aim, then generated 46 antibodies that measurably worked against a rare, aggressive childhood cancer, with no prior blueprint of the target. That's an unusually large batch of wins from a single design campaign.

The catch: these are lab-dish results. Nothing here has been tested in animals or people, and drug development takes years and often fails anyway. Amazon is also selling this as a service, so it has an interest in the hype. Still, if the approach holds up, it could mean cheaper drugs and shorter waits for patients who currently have few options.