Researching and Improving AI Agents for Biology
TxBench–Antibody Development: Benchmarking AI Agents on the Decisions That Take a Therapeutic Antibody from Lead Nomination to the First Human Dose
Preprint · TxBench–Antibody Development60 evals · 39.4% top scoreRead paper
Once a therapeutic antibody is nominated, its development depends on a series of decisions: whether the molecule is developable, whether a batch meets specification, how fast it clears, whether it will provoke anti-drug antibodies (ADA), what the toxicology allows and what dose is safe to give first. AI agents are increasingly used to support these decisions, but no benchmark tests whether they can make them from the underlying records. We introduce TxBench–Antibody Development, a verifiable benchmark of 60 evaluations spanning the seven stages from lead nomination to the first-in-human (FIH) dose. Each evaluation is built from patents, regulatory reviews, trial protocols or published studies, asks for a decision with the quantities that support it, and is graded deterministically. Across 34 model–harness configurations (21 models, three attempts each), the highest pass rate was 39.4% (Grok 4.7 / Grok Build), followed by Opus 5 / Claude Code (37.2%) and Sonnet 5.5 / PI (36.7%). Six models led at least one of the 20 categories, and no failure class had a mean pass rate above 35%.