BioGeometry releases GeoFlow V4: 34.6% average experimental hit rate across 15 de novo design tasks, about 85% above the previous generation
Synopsis
BioGeometry released GeoFlow V4, a new-generation drug discovery engine that reaches a project-level average experimental hit rate of 34.6% across 15 de novo design tasks spanning 13 disease-relevant targets, about 85% higher than the previous generation, with top affinity about 100-fold improved. The release designs small proteins, cyclic peptides, nanobodies and IgG antibodies de novo within one framework, and adds GeoFlow Code, an agentic design-orchestration system that turns multi-round design into executable, adjustable workflows. On a cross-species PD-1 IgG design task, GeoFlow Code's self-orchestrated pipeline reached a 4.23% overall pass rate versus 0.16% for the expert-planned pipeline.
Interpretation
BioGeometry has moved its drug discovery engine GeoFlow to V4, where de novo design no longer covers antibodies alone but supports small proteins, cyclic peptides, nanobodies and IgG antibodies within a single framework. The previous generation, GeoFlow V3, was the first in China to break through VHH de novo design, with design targets concentrated on antibodies; V4 extends this to four molecular formats. Across 15 de novo design projects spanning 13 disease-relevant targets, V4 reached a project-level average experimental hit rate of 34.6%, about 85% above the previous generation, with experimental validation using binding or functional assays suited to each target.
Beyond hit rate, the designed molecules reach picomolar affinity and cover multipass transmembrane targets whose accessible recognition surface is limited. Top affinity improved about 100-fold over the previous generation. Small proteins against PD-L1 and Nectin-4 reached 278 pM and 148 pM, a nanobody against BCMA reached 598 pM, an IgG against IL36R reached 458 pM, and an IgG against the pain-related peptide target PACAP reached as low as 15 pM; antibodies designed against the three multipass targets STEAP1, CCR8 and GCGR all showed cell-binding activity with apparent EC₅₀ at nanomolar or lower.
GeoFlow Code, released alongside, is a programming language and execution framework specialized for protein design that couples a large language model's planning, decision-making and execution with GeoFlow's structure understanding and generation, autonomously handling literature search, epitope selection, design orchestration and result analysis. V3's virtual-evolution pipeline iterated candidate optimization under a fixed strategy through multi-step reasoning; V4 lets the design strategy itself be explored and executed autonomously. On a cross-species PD-1 IgG design task, designing molecules within 48 hours that recognize human, monkey and mouse PD-1 and block PD-L1 binding, GeoFlow Code's self-orchestrated pipeline reached a 4.23% overall pass rate under a unified computational screening standard versus 0.16% for the expert-planned pipeline, with a sequence-diversity metric 19 times higher.
The designs were also tested for functional activity and early developability, not just binding signal. Binding is not function; for signaling targets such as GPCRs, binders may produce different downstream effects. Of 16 tested cyclic peptides designed against MC4R, 13 showed agonist activity in cellular cAMP assays, over 80%; among more than a hundred VHH and IgG molecules, 93% of tested VHHs and 96% of tested IgGs met preset expression thresholds, all 94 tested molecules passed dsDNA and insulin nonspecificity assays with a 99% BVP-ELISA pass rate, and 98% of tested VHHs and 99% of tested IgGs had Tm₁ above 60°C.
Perspective
The report covers BioGeometry's announced GeoFlow V4 and its companion GeoFlow Code: 15 de novo design tasks spanning 13 disease-relevant targets, plus the cross-species PD-1 design, MC4R cyclic-peptide functional assays and developability evaluation of more than a hundred VHH and IgG molecules. These results come from the company's published technical report and release materials and are aimed at researchers working on AI-assisted molecular design. The material does not state follow-up validation plans on more targets or larger project scales.
The hit-rate, affinity and developability figures are results announced by BioGeometry and no independent reproduction is reported; the GeoFlow Code versus expert comparison was run only on the single cross-species PD-1 IgG task, so whether the pass-rate and diversity gaps hold on other tasks is not yet clear.
