AIntibody: 29-institution Nature Biotech benchmark blind-tests 511 AI-designed antibodies
jiqizhixin · x · 2026-09-05
A 29-institution consortium published AIntibody in Nature Biotechnology—the first prospective, blinded benchmark for AI antibody design, inspired by CASP.
- Design: no peeking at answers and no back-testing on known data; predictions on truly unknown sequences validated by real experiments
- Three tasks: in silico affinity maturation from first-round sequencing data, ranking HCDR3 clusters by affinity, and designing sequences not included in the selection
- Closed loop: every prediction is anchored to experimental affinity and developability measurements
- Result: 86.1% of submitted sequences passed validation
More from Research
- After 10 Years Teaching ML Math, He Says You Only Need 20% of University Math — TivadarDanka · 2026-09-05
- Daniel Whiteson is hiring a postdoc at the intersection of HEP and ML — DanielWhiteson · 2026-09-05
- Open-source RAGnarok-AI launches human-annotation study: can you trust LLM judges? — Ok-Swim9349 · 2026-09-05
- Tom Leinster's 442-page book "Entropy and Diversity: The Axiomatic Approach" freely available on arXiv — FrnkNlsn · 2026-09-05
- 21K-Parameter Model Generates Business Emails, Trains Stably — SrijSriv211 · 2026-09-05
- AI-designed C8 protein vector cuts mRNA dose 10,000x vs lipid nanoparticles — bravo_abad · 2026-09-05