TxBench–Antibody Development: 60 verifiable evals across 7 drug-development stages
kenbwork · x · 2026-11-05
Following its Antibody Discovery Benchmark, the team introduces TxBench–Antibody Development, a verifiable benchmark of 60 evaluations across seven stages of therapeutic antibody development, from lead nomination to first-in-human dose.
Design:
- Evals are stratified by development stage, data source, decision type, and evidence source.
- Probes real scientific judgment: whether a manufacturing change alters drug quality despite an unchanged sequence, whether low animal exposure is a dosing artifact vs. clearance, and reconciling pharmacology and safety evidence for first human dosing.
Results:
- Pass rates fall from molecular to clinical stages: glycosylation and analytical QC reach 30.6% and 33.3%; toxicology and FIH dosing only 13%.
- Grok 4.7 is the overall leader; Opus 5.5 ranks fourth overall but leads six of seven stages and at least one of 20 columns.
- Two of the top ten models score zero on binding/potency assays (3%) and evidence integration (2%).
- Only Opus 5.5 passes the cytokine-release evaluation (17%), but misread average-vs-most-sensitive donor cytokines or EC50 as minimal effect level.
More from Research
- Stanford's James Zou on AI agents entering the wet lab, in Nature — james_y_zou · 2026-11-05
- Hamel Husain: Benchmarks are used like credit scores, so criticism is warranted — HamelHusain · 2026-11-05
- ArkenOS now runs on Rust, freeing room for more parallel agents — _AustinCalvert_ · 2026-11-05
- CoRL paper GUILD: actively guiding robot data collection toward useful auxiliary tasks — RobobertoMM · 2026-11-05
- TAMPAgent at CoRL: coding agents + TAMP planning for zero-shot long-horizon mobile manipulation — dhadfieldmenell · 2026-11-05
- STOC 2027: ~75% of 1,665 submissions report AI involvement in proofs and results — fortnow · 2026-11-05