MoleculeMind has published a new technical report on its MMDesign platform, which uses artificial intelligence to create proteins from scratch. The report presents the results of 19 design projects covering 18 different targets and two types of molecules: nanobodies (VHH) and macrocyclic peptides.
In 18 out of 19 projects, the platform was able to generate at least one molecule that was confirmed by laboratory experiments to bind to the target, corresponding to a project success rate of 94.7%.
The MMDesign workflow includes a fixed process: generation, filtering, and validation. First, tens of thousands of potential candidates are computationally generated, which are then assessed for structural reliability, sequence naturalness, and physicochemical evaluation. This stage narrows the selection for real laboratory testing down to 9–50 candidates, with a median value of 21.
Out of all 453 tested candidates, 139 were confirmed as binders, yielding a hit rate of 30.7%. Fifteen projects resulted in at least two binding agents, providing researchers with opportunities for comparison and optimization.
GPCR Trials Proved Most Challenging
The greatest difficulties were encountered with GPCR receptors, which are one of the most important families of drug targets but present a complex challenge for antibodies. This is because these receptors are located in the inner cell membrane, change shape, and have a limited surface area accessible from outside the cell. Nevertheless, MMDesign achieved confirmed binding molecules for all six tested GPCR targets: CCR7, CCR8, EP2, GPR68, NK3R, and A2AR.
Hit rates varied from 12.1% to 100%, and the best binder for each target reached nanomolar affinity; the best result for GPR68 was 12.1 nanomolar. Furthermore, all 30 tested candidates for NK3R showed binding, as did 10 out of 11 for GPR68. For five GPCR targets for which no cataloged therapeutic antibodies existed, 53 out of 89 developed constructs demonstrated binding.
The report also examines result reproducibility. A second round of design on six targets that already had binding molecules improved the hit rate, best affinity, or both parameters in five cases. For example, the hit rate for PDGFR increased from 4.0% to 21.1%, and the best affinity improved from 150 to 37 nanomolar. Regarding macrocyclic peptides, 16 out of 21 MDM2 candidates and 11 out of 28 PD-L1 candidates showed binding. Some candidates also demonstrated good expression profiles, minimal aggregation, and low non-specific binding in early development tests.
The basis for these developments is MoleculeMind's full atomic structure prediction model—MMFold. In the FoldBench antigen-antibody benchmark, covering 172 interfaces, the MMFold model achieved a top-1 high-quality prediction success rate of 39.0%, roughly double the score of AlphaFold 3, which was 19.8%. The full report is available on Zenodo.