MoleculeMind reports 18 of 19 design projects hit, with all six GPCR targets bound

MoleculeMind's technical report describes a fixed generate-screen-verify loop: tens of thousands of candidate molecules are generated in silico, filtered on structural reliability, sequence naturalness and statistical-physics scoring down to a few dozen, then handed to the wet lab.

The evaluation spans two modalities, VHH nanobodies and macrocyclic peptides, across 18 targets and 19 design projects. Each project tested only 9 to 50 candidates, with a median of 21, and 18 projects produced experimentally confirmed binders, a 94.7% project success rate.

At the candidate level the numbers are lower but disclosed: 139 of 453 tested designs bound, a 30.7% hit rate, and 15 projects yielded at least two binders, giving teams room to compare and optimize instead of settling for the first hit.

The GPCR panel is the harder test. All six targets, CCR7, CCR8, EP2, GPR68, NK3R and A2AR, produced confirmed binders with best affinities in the nanomolar range; GPR68 reached a best KD of 12.1 nanomolar, NK3R bound with all 30 candidates tested, and GPR68 with 10 of 11.

The no-precedent scenario is the strongest evidence for de novo design. On five GPCRs with no catalogued therapeutic antibody precedent in Thera-SAbDab, 53 of 89 tested designs bound, a 59.6% hit rate, which is performance without an answer key.

Reuse shows up twice. Re-running design on six previously solved targets improved five of them: PDGFR's hit rate rose from 4.0% to 21.1% and its best KD from 150 to 37 nanomolar. On macrocyclic peptides, the MDM2 project hit 16 of 21 candidates and PD-L1 hit 11 of 28.

Developability was folded into screening as well: some VHH-Fc candidates expressed well in 10-milliliter CHO cultures, purified as mostly non-aggregated main peaks, and showed low binding to insulin and double-stranded DNA as off-target antigens; comparisons against PDB, OAS and SAbDab found substantial divergence in antigen-recognition regions, including novel CDR3 conformations.

The structural layer underneath is MMFold, which the report says reached 39.0% top-1 high-quality success across all 172 antibody-antigen interfaces in FoldBench, roughly double AlphaFold 3's 19.8%, with 70.9% top-1 and 76.7% top-5 overall success.

All of these figures come from the company's own report, have not been independently reproduced, and do not yet include animal or clinical data. If the generate-screen-verify loop holds across more targets and modalities, AI drug design moves from one-off demonstrations to infrastructure that pharma can call on repeatedly.