BioGeometry's GeoFlowV4 hits 34.6% wet-lab hit rate across 13 drug targets, up 85%
机器之心 · wechat · 2026-11-20
AI drug discovery company BioGeometry released GeoFlowV4, a next-generation drug R&D engine, along with GeoFlowCode, an agent-based orchestration framework for protein design. The headline result: across 13 disease targets and 15 de novo design tasks, the average wet-lab hit rate reached 34.6%, an 85% improvement over the previous generation, with positive validations across small proteins, cyclic peptides, nanobodies, and full-length IgG, and up to 100x better binding affinity.
On benchmarks, GeoFlowV4 achieved a 75.4% Top-1 success rate on the hard FoldBench-AB antibody-antigen structure prediction test, rising to 78.8% with 100 inference samples. Its confidence scoring enriches binders 6x among top-5% ranked candidates. GeoFlowCode autonomously plans multi-round design campaigns: on a cross-species PD-1 IgG task, its pass rate of 4.23% was 26x that of expert-planned pipelines.
Developability data also held up: 93-96% of tested antibodies met expression targets, 98-99% met thermal stability targets, and 99% passed non-specific binding assays. Designs against multi-pass GPCR targets (STEAP1, CCR8, GCGR) showed cell-surface binding activity. The piece frames a industry-wide shift from 'can it bind' toward 'is it developable as a drug.'
More from Apps
- Tesla says FSD is now approved in Italy, rollout begins soon — bendomingueztv · 2026-11-20
- HK AI security startup IRONWARD gets $10k OVHcloud credits to build agent-run security ops — AIFlow_ML · 2026-11-20
- Deskworlds: open-source living 3D desktop worlds for Mac, built entirely with code — chaseleantj · 2026-11-20
- X Algorithm Now Silently Drops Mentions Containing Muted Words, Code Shows — ThePeterMick · 2026-11-20
- Designer lays out 5 ways to use AI and make money in 2026 — Tegadesigns · 2026-11-20
- Open-source MCP server gives your ChatGPT or Claude an A2A address so AIs can haggle with each other — IDforOpus · 2026-11-20