AI-native bioengineering infrastructure company MoleculeMind has reported experimentally confirmed binders in 18 of 19 design campaigns, representing a 94.7% campaign-level success rate, with just nine to 50 candidates tested per campaign, according to a technical report released in October 2026. Its platform, MMDesign, addresses a practical challenge in AI-driven drug discovery by turning computational designs into molecules that bind their intended targets while keeping the number of candidates sent for laboratory testing small.
The results span 18 distinct biological targets and two molecular modalities, nanobodies and macrocyclic peptides. Across the campaigns, 139 of 453 tested candidates showed confirmed binding, an overall candidate-level hit rate of 30.7%, with a median of 21 candidates tested per campaign. Fifteen campaigns produced at least two binders, giving researchers more than one starting point for further characterization and potential optimization.
Results for G protein-coupled receptors (GPCRs) offer a closer look at performance on targets whose membrane-associated structures and conformational flexibility can complicate antibody discovery. MMDesign yielded binders against all six GPCR targets evaluated, with the best-performing binder for each showing an apparent binding affinity in the nanomolar range under the reported biolayer interferometry assay conditions. All 30 NK3R candidates tested showed binding, while 10 of 11 GPR68 candidates were confirmed as binders, with the best-performing GPR68 candidate reaching an apparent dissociation constant, or KD, of 12.1 nanomolar.
The report also examines whether a successful campaign can be improved in a subsequent design round. Of six previously successful targets revisited, five showed gains in hit rate, best measured affinity or both. For platelet-derived growth factor receptor, or PDGFR, the hit rate rose from 4.0% to 21.1%, while the best measured KD fell from 150 to 37 nanomolar, an approximately fourfold improvement in binding affinity.
The work extends beyond nanobodies to macrocyclic peptides, broadening the molecular formats evaluated. In the MDM2 campaign, 16 of 21 candidates showed binding, a 76.2% hit rate, with a best measured apparent KD of 13.7 micromolar. For PD-L1, 11 of 28 candidates were confirmed as binders, with a best measured apparent KD of 40.2 micromolar.
Supporting candidate selection is MMFold, MoleculeMind’s all-atom structure prediction model, used alongside sequence-based and physics-based evaluation methods. In the company’s reported FoldBench evaluation across 172 antibody-antigen interfaces, MMFold achieved a 39.0% top-1 high-quality structure prediction success rate, compared with the reported AlphaFold 3 baseline of 19.8%. These figures describe a specific benchmark, rather than performance across all structure prediction tasks.
The study also reports encouraging preliminary developability results for selected nanobody-based candidates, covering expression, monomeric content and nonspecific binding to surrogate antigens. MoleculeMind identifies affinity optimization, more complex targets and molecular formats, and broader developability assessment as priorities for further work.
The results suggest a way for computational design to complement established discovery methods by concentrating laboratory testing on smaller candidate panels while providing starting points across different targets and molecular formats.