LLM-designed antibodies hit 38/80 binders in lab tests, 85% success on RAGE
DeryaTR_ · x · 2026-09-09
A team built opendde-harness, a workflow that uses LLM reasoning to design antibodies and then validated the designs experimentally: 80 synthesized designs across three targets produced 38 experimentally detected binders.
Success rates by target–format cohort:
- RAGE: 17/20 (85%)
- TfR1: 11/20 (55%)
- CXCR4 VHH: 6/20 (30%)
- CXCR4 mAb: 4/20 (20%)
Notably, CXCR4 is a 7-transmembrane protein notoriously hard to target with synthetic antibodies due to its dynamic conformations, yet 10 designs still bound. A commenter who co-discovered CXCR4 as an HIV co-receptor 30 years ago called it a super impactful BioAI model. Example workflow: ask the LLM to design a VHH against CRLF2, verifying target and epitope, keeping the framework fixed, and designing CDRs with a visible plan.
More from Research
- Cognition's Scott Wu claims AI solved Navier-Stokes 'in about a week' — ivan_bezdomny · 2026-09-09
- vivo AI Lab's CAPO-SLT fixes RL drift in sign language translation with confidence-aware clipping — 量子位 · 2026-09-09
- Algorithm Reproduces All Known Transversal Gates on Stabiliser Codes at 69-100% Success — maier_ak · 2026-09-09
- Variational Pipeline Learns Target Logical Gates from Encoding Unitary Alone — maier_ak · 2026-09-09
- Delip Rao: AI peer review will die — abstract bidding lets bad actors steal submissions — deliprao · 2026-09-09
- Machine civs lack mortality's regularization — 'like training MoE without load balancing' — kellerjordan0 · 2026-09-09