MMFold

A co-folding model places an interface in large part by reading evolutionary signal: residues that covary across a multiple sequence alignment are residues that touch. Antibody–antigen recognition supplies none of it. An antibody and its antigen share no evolutionary history, the paratope sits on hypervariable CDR loops that differ between clones of the same germline, and the epitope is picked by somatic selection rather than conserved down a family. So the interface class that therapeutic biologics engineering depends on is precisely the one where general-purpose all-atom predictors, otherwise near-experimental on obligate complexes, land closer to a coin flip.

MMFold is MoleculeMind's answer to that gap: a proprietary all-atom structure prediction model inspired by the AlphaFold 3 architecture and trained on Protein Data Bank entries released before 30 September 2021, with the company reporting enhanced antibody–antigen modeling capability from that training. It was documented in a June 2026 technical report from the Shanghai protein-design company founded by computational biologist Jinbo Xu, alongside the nanobody design platform MMDesign, whose hallucination loop optimizes designs against MMFold's confidence landscape without ever updating its weights.

The model is closed. No weights, code, checkpoint, or public API have been released, and the only route to it is as a feature of MoleculeOS, MoleculeMind's login-gated platform. Every performance figure below is the company's own.

#Key Features

• Antibody–antigen as the design target: Rather than treating immune complexes as one more category in a general co-folding benchmark, MMFold's reported training emphasis is on exactly this interface class, and its published evaluation is confined to it.
• Accuracy that scales with sampling: Success rises from 68.6% at Top-1 to 75.6% at Top-5, and the report notes that with a few hundred seeds the high-quality DockQ rate approaches 50% at Top-5 — compute buys accuracy at inference rather than requiring a larger model.
• A leakage-free evaluation window: The 2021 training cutoff predates the FoldBench release window entirely. The report excludes Boltz-2 from its comparison for failing that condition, a constraint MMFold satisfies by construction.
• A frozen oracle for generative design: Because it is applied unchanged to arbitrary new targets, MMFold serves as the differentiable objective a design loop can hallucinate against, and its confidence metrics feed the filtering cascade that ranks candidates.

#Technical Details

Architectural disclosure stops at "inspired by the AlphaFold 3 architecture." Parameter count, training-corpus size, compute, and the specifics of how antibody–antigen emphasis was applied are not published. What is reported is the benchmark: all 172 antibody–antigen interfaces (113 PDB entries) of FoldBench, scored by DockQ. MMFold reached 68.6% acceptable-or-better at Top-1, of which 61.6% was medium-quality or better and 34.9% high-quality; at Top-5 those figures were 75.6%, 65.7% and 39.5%. Under the same protocol AlphaFold 3 scored 47.9%, Boltz-1 33.5%, HelixFold3 28.4% and Chai-1 23.6%; the Protenix v1 (47.0%) and ESMFold2 (55.0%) figures were taken from the ESMFold2 paper rather than rerun in house. Protenix-v2's reported 65% was measured on a 160-interface subset and was left out as not directly comparable. Case studies show the gap concretely: on PDB entries 8AHN, 8PIH and 8VYL, AlphaFold 3 returned DockQ scores of 0.020, 0.273 and 0.009 against MMFold's 0.869, 0.963 and 0.749.

#Applications

MMFold is aimed at antibody and nanobody engineering, where a predicted complex is the input to epitope verification, interface analysis, and candidate triage. Inside MoleculeMind's own pipeline it did both jobs: it generated the 363 designed VHH structures across 11 target systems that the report analyzed for structural novelty against SAbDab, and its confidence output was one of four filtering tiers that cut tens of thousands of generated candidates to the 14–50 per target sent to the bench. External users reach it only through MoleculeOS, which the company opened to industry customers in July 2026.

#Impact

MMFold is one of the few structure predictors to make antibody–antigen accuracy its headline claim rather than a subsection, and the margin it reports over AlphaFold 3 on FoldBench is large enough to matter for design workflows that spend that accuracy directly on hit rates. The evidence base is thin in ways worth stating plainly. The sole source is a company technical report, not a preprint or a peer-reviewed paper; MoleculeMind ran the benchmark itself rather than submitting to its authors, and two of the six baselines were transcribed from a third paper rather than evaluated under the common protocol. One anchor does hold: the AlphaFold 3 baseline of 47.9% matches the figure FoldBench's own authors published. With no released weights and no open endpoint, none of it can be independently reproduced, and the model's behavior outside antibody–antigen complexes — on nucleic acids or ligands, as an all-atom AlphaFold 3-class model would be expected to handle — is undocumented.