MoleculeMind MMDesign Technical Report: 18 of 19 Projects Succeeded, All 6 GPCR Targets Hit

QbitAI | WeChat Official Account QbitAI

A technical report recently released by MoleculeMind, however, offers a somewhat different answer.

This latest technical report on its AI de novo protein design platform MMDesign covers 18 different targets and a total of 19 design projects across two molecular modalities, nanobodies (VHH) and macrocyclic peptides:

Together, these numbers point to MMDesign having moved from "technical breakthrough" into the "systematic validation" stage: the same "generate-screen-validate" workflow can, across projects spanning targets and modalities, repeatedly pass wet-lab validation with a small number of candidate molecules.

MMDesign, which has undergone three iterations within half a year, seeks to turn "efficient hits" into a repeatable capability.

After generation, can fewer candidate molecules be sent to the lab for testing, and can the hit rate be higher? This is efficiency.

Can the affinity, developability metrics and novelty of candidate molecules be better? This is quality.

Another point is equally critical: the generation process does not rely on any known binding molecule as a reference for modification, but designs molecules directly from functional requirements, and can deliver answers even for targets with no precedent to follow. This is precisely the essential value of de novo design as distinct from "modifying known molecules."

But generating sequences is only the starting point. From tens of thousands of candidates, precisely picking the few dozen worth taking into the lab is the real watershed.

MMDesign fixes this process into a unified "generate-screen-validate" workflow:

First, tens of thousands of candidates are generated in computational space; then, through multi-layer evaluation based on structural reliability, sequence naturalness and statistical physics, the candidate set for a single project is compressed to just 9-50, with a median of only 21. Of the 19 projects, 18 yielded bindable molecules, for a project success rate of 94.7%; of all 453 tested candidates, 139 were confirmed to bind, for an overall candidate hit rate of 30.7%.

Here there are two yardsticks, different in scope but both important: 94.7% measures "whether a project can find at least one binding molecule," while 30.7% measures "how many of the tested candidates actually hit." For de novo design, both figures are already quite impressive globally.

15 projects obtained at least two binding molecules, accounting for nearly 80% of all projects, meaning R&D teams not only "have something to work with" but can also compare and optimize among multiple candidates.

For pharmaceutical companies, excluding low-value candidates as early as possible on the computational side allows experimental resources to be concentrated on high-potential molecules. This is the real value of shifting from "mass screening" to "precise design."

If 94.7% demonstrates efficiency, the GPCR series directly tests the technology's mettle.

GPCRs are one of the most important drug target families in the pharmaceutical industry, but designing antibodies against them has long been difficult.

On the one hand, as membrane proteins, GPCRs have complex structures and variable conformations; on the other hand, the regions exposed outside the cell and available for antibody binding are very limited. These characteristics make GPCRs a recognized "hard nut to crack" in the field of de novo antibody design.

The report shows that MMDesign carried out de novo design against six GPCR targets—CCR7, CCR8, EP2, GPR68, NK3R and A2AR—and all successfully yielded experimentally confirmed binding molecules.

Under the BLI assay conditions used in the report, candidate hit rates for the various targets ranged from 12.1% to 100%, and the affinity of the best binding molecule for all six targets reached the nanomolar level, with the best KD for GPR68 as low as 12.1 nanomolar.

Two sets of results stand out in particular: all 30 tested candidates for NK3R were confirmed to bind, a hit rate of 100%; and 10 of 11 candidates for GPR68 hit.

Further evidence comes from "no precedent" scenarios:

On five GPCR targets for which a search of the Thera-SAbDab database (PL132) found no catalogued precedent therapeutic antibodies, 53 of 89 tested designs bound, for a candidate hit rate of 59.6%.

A 59.6% candidate hit rate shows that even without an "answer key," the platform can still open up new discovery paths on complex membrane protein targets, offering the possibility of breaking through traditional R&D bottlenecks.

What pharmaceutical companies and platforms truly value in long-term collaboration is "whether it can be used repeatedly": after an initial design, can it still be optimized? If the task changes, can it still succeed?

This report gives an affirmative answer: MMDesign conducted another round of design on six targets that already had binding molecules, and five of them achieved further improvements in hit rate, best affinity, or both.

Taking PDGFR as an example, the candidate hit rate rose from 4.0% to 21.1%, and the best KD was optimized from 150 nanomolar to 37 nanomolar, corresponding to a 4-fold affinity improvement. For the same R&D problem, another round can still deliver better molecules.

The platform's scope of application also extends to different molecular modalities: macrocyclic peptide design results show that for the MDM2 project, 16 of 21 candidates hit, a hit rate of 76.2%, with a best KD of 13.7 micromolar; for the PD-L1 project, 11 of 28 candidates hit, with a best KD of 40.2 micromolar.

Compared with nanobodies, macrocyclic peptide conformations are constrained by the cyclic backbone, so the design challenges differ, but the approach works just as well.

This "generate-screen-validate" capability is being reused for different design tasks.

Cross-round design optimization and cross-modality workflow reuse together form two sides of systematic validation:

The platform's capability is moving from "single-point success" toward "systematic output."

For drug discovery, "being able to bind" counts as getting a ticket in, but a truly valuable early hit molecule still has to pass several later "checkpoints":

Can it be expressed stably, will it aggregate, and does it bind irrelevant proteins nonspecifically?

The report shows that MMDesign has already moved such issues forward into the design evaluation stage: some VHH-Fc candidates showed good expression levels in CHO cell expression experiments at a 10 mL culture scale; many purified candidates were dominated by a non-aggregated main peak component, with few signs of aggregation; and the tested candidate subset also showed low nonspecific reactivity against two surrogate antigens, insulin and double-stranded DNA.

While pursuing binding activity, the platform has already taken into account the key properties required for molecules to enter subsequent drug development.

Sequence novelty is even more direct evidence of "design without a reference answer."

MMDesign's generation process does not rely on retrieving known binding molecules from antibody sequence databases as starting points for modification; comparisons with databases such as PDB, OAS, and SAbDab show that designed molecules for multiple targets differ significantly from known antibodies in the antigen recognition region, and the key binding loop region CDR3 of some molecules presents novel conformations.

The candidates produced by the platform are not "copies" of known molecules, but new sequences generated around functional requirements, directly targeting problems that are difficult for traditional R&D to overcome, and also offering the possibility of bypassing existing patent layouts and expanding entirely new candidate space.

For de novo design to "select accurately," structure prediction must first "see accurately."

MMFold, the all-atom structure prediction model independently developed by MoleculeMind, is the structural foundation behind MMDesign.

On FoldBench, the internationally authoritative benchmark for antibody-antigen complex structure prediction, MMFold achieved a Top-1 high-quality structure prediction success rate of 39.0% across evaluations of all 172 interfaces, about twice the 19.8% comparison result of AlphaFold 3;

The Top-1 overall success rate reached 70.9%, and the Top-5 overall success rate further reached 76.7%, among which the high-quality prediction success rate was 44.2%.

On the subset of 160 shared interfaces corresponding to the public results of OpenDDE and Protenix-v2, MMFold's Top-1 high-quality success rate reached 39.4%, higher than the 33.0% reported by OpenDDE and the 28.2% reported by Protenix-v2; the Top-5 high-quality success rate further reached 45.0%, with an overall success rate of 78.8%.

The value of structure prediction lies not in the ranking itself, but in giving candidate molecule evaluation and prioritization a more solid computational basis.

With more accurate structures, molecular designers can more quickly lock in, from a vast computational space, the few candidates worth experimenting on.

The multimodal protein foundation large model NewOrigin (Darwin) provides underlying capabilities, MMDesign handles de novo generation and candidate prioritization, MMFold focuses on structure prediction, QuantaMind performs dynamic simulation, and the AI-native R&D operating system MoleculeOS connects model tools and R&D processes into a closed loop of "generation-computation-experiment-iteration."

MoleculeMind founder Xu Jinbo said:

The important significance of this expanded evaluation lies in gradually transforming one-off design success into repeatable, scalable R&D capability. The long-term value of AI biopharmaceutical R&D is not just performance breakthroughs of a single model, but rather, through the coordination of proprietary underlying models with generation, evaluation, and experimental validation, making precise molecular design a technical infrastructure capable of continuously solving different R&D problems.

From single-point hits to systematic validation across targets and modalities, from underlying model breakthroughs to reusable R&D workflows, MMDesign's latest report advances the question of "whether AI can design molecules" to "whether AI can stably and reproducibly produce usable molecules."

The design efficiency of few candidates and stable hits; molecular quality that balances affinity, developability, and novelty; and the breakthrough value of not relying on "reference answers," directly producing molecules from functional requirements, and successfully tackling highly difficult targets. Together, the three support "reusability," enabling AI molecular design to be continuously used by industry to solve real R&D needs.

*This article is published by QbitAI with authorization; the views are solely those of the original author.