分子之心's MMDesign Gets Binders in 18 of 19 Antibody and Peptide Design Projects, All 6 GPCR Targets Hit

分子之心 reports that its AI protein de novo design platform MMDesign produced experimentally confirmed binders in 18 of 19 design projects, a 94.7% project success rate. The company published the figures in a technical report covering 18 targets and two molecule types: nanobodies (VHH) and macrocyclic peptides. The details come from an article by Chinese outlet QbitAI (量子位), published in Chinese on 2026-10-28; the report itself is hosted on Zenodo.

Screening efficiency

MMDesign generates tens of thousands of candidates computationally, then narrows them through structural reliability, sequence naturalness and statistical-physics-based evaluation. Each project tested 9 to 50 candidates, with a median of 21.

Across all projects, 139 of 453 tested candidates were confirmed binders, an overall candidate hit rate of 30.7%. Fifteen projects yielded at least two binders. The company says the platform went through three iterations in half a year.

GPCR results

The report covers six GPCR targets: CCR7, CCR8, EP2, GPR68, NK3R and A2AR. All six produced experimentally confirmed binders. Under the report's BLI assay conditions, per-target candidate hit rates ranged from 12.1% to 100%, and the best binder for every target reached nanomolar affinity.

• NK3R: all 30 tested candidates bound (100%).
• GPR68: 10 of 11 candidates bound; best KD was 12.1 nM.
• Five GPCR targets with no cataloged therapeutic antibody precedent in Thera-SAbDab (PL132): 53 of 89 tested designs bound, a 59.6% hit rate.

Repeat rounds and macrocyclic peptides

For six targets that already had binders, MMDesign ran another design round. Five improved in hit rate, best affinity, or both. On PDGFR, the hit rate rose from 4.0% to 21.1%, and best KD improved from 150 nM to 37 nM, which the report describes as a 4x affinity gain.

For macrocyclic peptides:

• MDM2: 16 of 21 candidates bound (76.2%); best KD 13.7 μM.
• PD-L1: 11 of 28 candidates bound; best KD 40.2 μM.

Developability and novelty

The report says some VHH-Fc candidates expressed well in 10 mL CHO cell cultures. Many purified candidates were mostly a non-aggregated main peak. A tested subset showed low non-specific reactivity to two surrogate antigens, insulin and double-stranded DNA.

On novelty, the generation process does not retrieve known binders from antibody sequence databases as starting points. Comparisons against PDB, OAS and SAbDab show that designs for several targets differ markedly from known antibodies in the antigen-recognition region. Some show novel CDR3 conformations.

MMFold structure prediction

MMDesign relies on MMFold, the company's all-atom structure prediction model. On the FoldBench antibody-antigen benchmark (172 interfaces), MMFold reached 39.0% Top-1 high-quality success, against 19.8% for AlphaFold 3 as cited in the report. Top-1 overall success was 70.9%, Top-5 overall success was 76.7%, and Top-5 high-quality success was 44.2%.

On a 160-interface subset shared with OpenDDE and Protenix-v2 public results, MMFold's Top-1 high-quality success was 39.4%, against 33.0% for OpenDDE and 28.2% for Protenix-v2. Its Top-5 high-quality success was 45.0% and overall success 78.8%.

The company's stack also includes the multimodal protein foundation model NewOrigin (达尔文), QuantaMind for dynamic simulation, and the AI-native R&D operating system MoleculeOS.

What this means

The report gives two separate metrics: 94.7% measures whether a project found at least one binder, and 30.7% measures how many tested candidates bound. The first is the better indicator of project-level reliability; the second reflects per-candidate efficiency. All figures are company-reported results from its own technical report.

The GPCR and no-precedent results matter most because they test design without a known binder to modify. The repeat-round gains, such as PDGFR, show iterative improvement on the same target. Macrocyclic peptide best affinities were in the micromolar range, well behind the nanomolar nanobody results. Founder 许锦波 (Xu Jinbo) says the goal is to turn one-off design successes into a repeatable R&D capability.