MMDesign

Conventional antibody discovery interrogates 10⁶ to 10⁹ variants — an immunized animal's repertoire, a display library — and then asks which survivors happen to bind where you wanted. Epitope specificity is something you screen for afterwards rather than something you request. MMDesign inverts that arithmetic: from a target protein sequence and the epitope residues to engage, it generates tens of thousands of VHH candidates computationally and hands the wet lab 14 to 50.

The generative step is hallucination. Rather than sampling from a learned model of antibody sequence space, MMDesign co-optimizes a nanobody's CDR sequence and structure against the confidence landscape of a frozen all-atom structure predictor — MMFold, MoleculeMind's proprietary antibody–antigen complex model — together with a protein language model. No weights are updated and no per-target training is required; the optimization moves the design, not the model. That puts it in the same family as Germinal and BoltzDesign1, which likewise steer a frozen predictor rather than train a dedicated generator; the report positions MMDesign alongside both and the purpose-built generative system Chai-2. Because generation never consults an antibody sequence database, designs cannot by construction be copies of deposited or patented binders.

MMDesign was released in June 2026 by MoleculeMind (分子之心), a Shanghai protein-design company founded by computational biologist Jinbo Xu, and documented in a technical report. The name is shared with an unrelated inverse-folding framework from Westlake University; this entry covers the nanobody discovery platform.

#Key Features

• Epitope-directed from sequence alone: Inputs are a target protein and the hotspot residues a binder should contact — no antigen structure, no starting antibody, no target-specific dataset.
• Hallucination over a frozen oracle: CDR sequence and backbone are optimized jointly against the predictor's own confidence signal, so search is guided by predicted complex quality, not similarity to known binders.
• Four-stage computational filter: Structural confidence, reproducibility across random seeds and prediction models, sequence naturalness under a protein language model, and an in-house physics-aware interface score are layered. No single metric sufficed alone — ipTM and pLDDT correlated poorly with measured affinity.
• Developability assessed alongside binding: Candidates were expressed as VHH-Fc constructs in CHO cells, profiled by SEC-HPLC, and counter-screened against human serum albumin, insulin and dsDNA for non-specific reactivity.

#Technical Details

MMFold, the predictor at the center of the loop, is an all-atom model inspired by the AlphaFold 3 architecture, trained on Protein Data Bank entries released before 30 September 2021 with enhanced antibody–antigen modeling capability. On FoldBench's 172 antibody–antigen interfaces it reached a DockQ acceptable-or-better rate of 68.6% for Top-1 predictions against 47.9% for AlphaFold 3 under the same protocol, rising to 75.6% at Top-5. The generative core the loop optimizes against is described as a general-purpose all-atom structure predictor not specifically fine-tuned on antibody–antigen complexes. Parameter counts, training-corpus size, and the protein language model's identity are not disclosed.

Across 11 therapeutic targets, testing 14–50 candidates each yielded confirmed binders for 10 (90.9%) by biolayer interferometry. PD-L1 gave the highest hit rate at 26 of 30 with a best K_D of 7.2 nM; TNFα, a shallow homotrimeric cytokine that has resisted prior low-throughput de novo campaigns, gave 7 of 14 with an apparent K_D of 51 pM reflecting avidity against the trimer. InsulinR (5/25, 25 nM), BHRF1 (5/35, 282 nM), IL-2 (5/50, 138 nM) and IL-3 (2/25) yielded multiple binders; PDGFR, IL-7Rα, IL-20 and PD-1 one apiece; CD7 none. Nine of the eleven targets have no VHH complex in the PDB; framework regions held canonical VHH geometry (Cα-RMSD below 0.7 Å) while CDR3 conformations diverged from the experimental database. Where references exist the designs converged on the right site anyway — the PD-1 binder lands on the natural PD-1:PD-L1 interface.

#Applications

MMDesign targets therapeutic nanobody discovery where experimental throughput, not computational capacity, is the constraint — a lab that can characterize a few dozen constructs but not run a display campaign. The epitope-hotspot input suits cases where a specific functional site must be blocked, and the evaluated panel spans immunology, oncology, metabolism and infectious disease. It is offered for collaborative campaigns with pharmaceutical and biotechnology partners; MMFold is available separately through MoleculeOS.

#Impact

MMDesign demonstrates that epitope-targeted generative nanobody design can run at a scale a normal biologics lab can absorb, across a panel broader than the handful of demonstration antigens typical of the field. The caveats are substantial. No weights, code, or public inference endpoint have been released and the platform is reachable only through partnership, so the results cannot be independently reproduced. Affinities are variable, clustering in the nanomolar regime rather than consistently sub-nanomolar; the geometries rest on predicted structures, with no experimental structure of a designed complex; and stability, immunogenicity and pharmacokinetics were outside the study's scope. The source is a company technical report rather than a peer-reviewed paper; its numbers differ from the launch announcement, which adds a twelfth campaign against the GPCR CCR7 absent from the report.