Skip to content
MMDesign Achieves 94.7% Success Rate Across 19 Targets with Only Dozens of Experimental Tests

MMDesign Achieves 94.7% Success Rate Across 19 Targets with Only Dozens of Experimental Tests

Molecule Heart's MMDesign platform demonstrates a reproducible AI de novo design workflow achieving 94.7% project success rate across 19 targets including six GPCRs, with only 9-50 candidates tested per project, and shows iterative improvement and macrocyclic peptide extension backed by the high-performing MMFold structure prediction model.

Machine Heart
Machine Heart
Machine Heart

From Single Breakthroughs to Systematic Validation

Molecule Heart's latest technical report for its AI de novo design platform MMDesign presents a shift from generating thousands of sequences to delivering experimentally validated hits with minimal screening. Across 19 design projects covering 18 targets and two modalities (nanobodies/VHH and macrocyclic peptides), 18 projects yielded at least one experimentally confirmed binder — a 94.7% project success rate . Each project required testing only 9–50 candidates (median 21) , and 453 total tested candidates produced 139 confirmed binders, a 30.7% candidate hit rate . Fifteen projects provided at least two binders, preserving options for downstream optimization.

GPCR Targets: All Six Hit, Including Novel Targets

GPCRs are notoriously difficult membrane proteins. MMDesign tested six GPCR targets — CCR7, CCR8, EP2, GPR68, NK3R, and A2AR — and all six produced confirmed binders . Two standout results: NK3R achieved 100% candidate hit rate (30/30) , and GPR68 hit 10 of 11 candidates with a best apparent KD of 12.1 nM . Combined, 41 candidates yielded 40 binders. Moreover, on five GPCRs with no previously cataloged therapeutic antibodies in Thera-SAbDab, 89 tested candidates gave 53 binders — a 59.6% candidate hit rate — demonstrating applicability to targets lacking prior structural or antibody knowledge.

Iterative Improvement on Repeat Targets

For six targets previously addressed by MMDesign, a new design round improved outcomes in five projects . The PDGFR project exemplifies this: candidate hit rate rose from 4.0% to 21.1% (5.2×) , and best KD improved from 150 nM to 37 nM (4×) . This shows the platform can iteratively optimize the same target rather than relying on a single-shot design.

Modality Expansion to Macrocyclic Peptides

Beyond nanobodies, two macrocyclic peptide projects succeeded: MDM2 (16/21 binders, best KD 13.7 μM) and PD-L1 (11/28 binders, best KD 40.2 μM) . This confirms the workflow extends to chemically distinct modalities.

Developability and Sequence Novelty

Many MMDesign VHH-Fc candidates expressed stably in CHO cells, remained monomeric after purification, showed low aggregation, and exhibited low non-specific binding against insulin and dsDNA — positive early developability signals. Sequence comparisons against PDB, OAS, and SAbDab revealed novelty in both full sequences and binding regions, enabling differentiated IP landscapes.

Underlying Model: MMFold Outperforms AlphaFold 3 on Antibody–Antigen Interfaces

MMDesign's precision relies on Molecule Heart's proprietary structure prediction model MMFold . On the FoldBench benchmark of 172 antibody–antigen interfaces , MMFold Top-1 achieved 70.9% overall success and 39.0% high-quality success — approximately 2× AlphaFold 3's high-quality rate . On a shared subset of 160 interfaces, MMFold's high-quality rate was 39.4% , surpassing OpenDDE (33.0%) and Protenix-v2 (28.2%) . Accurate interface prediction directly feeds candidate selection in the design loop.

FoldBench all 172 antibody–antigen interface structure prediction results
FoldBench all 172 antibody–antigen interface structure prediction results

FoldBench all 172 antibody–antigen interface structure prediction results

FoldBench 160 common interface subset structure prediction results
FoldBench 160 common interface subset structure prediction results

FoldBench 160 common interface subset structure prediction results

Platform Ecosystem and Paradigm Shift

Molecule Heart's stack comprises NewOrigin (foundation model), MMDesign (de novo generation and candidate selection), MMFold (structure prediction), QuantaMind (molecular dynamics), and MoleculeOS (workflow orchestration). The report argues the industry is moving from "high-throughput screening" to a "define needs → precise design → low-throughput validation → continuous optimization" paradigm, and MMDesign's cross-target, cross-modality, iterative results provide empirical support for this transition.

Source: Molecule Heart MMDesign latest technical report (https://zenodo.org/records/23321531).
Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

drug discoverystructure predictionAI de novo designGPCRmacrocyclic peptideMMDesignMMFoldnanobody
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.