Published October 12, 2026
| Version v1
Report
Open
MMDesign: Robust De Novo Binder Discovery Across Diverse Targets and Modalities
Authors/Creators
Description
We present an expanded MMDesign, a platform for robust de novo binder discovery across diverse target classes and molecular modalities. Combining generative sequence–structure optimization with structural, sequence-based, and statistical-physics-based prioritization, MMDesign applies a shared generate–filter–validate workflow to non-GPCR protein targets, challenging GPCR membrane proteins, and both VHH and macrocyclic binders. This breadth makes the platform useful across discovery programs with different target biology and binder formats.
The consolidated evaluation spans 19 campaigns on 18 distinct targets, with confirmed binders in 18 of 19 campaigns (94.7%) while experimentally testing 9–50 candidates per campaign. VHH discovery succeeded in 16/17 campaigns, including all six tested GPCRs, with 30/30 hits for NK3R and 10/11 for GPR68. Re-evaluation of six previously challenging targets improved hit rate, affinity, or both for five targets, demonstrating repeatable improvement across design rounds. Macrocyclic binders were recovered for both MDM2 and PD-L1, extending the workflow beyond antibody scaffolds. Together, these results support robustness across the evaluated target classes and practical versatility across two binder modalities.
MMFold provides the platform’s structural modeling foundation, achieving 39.4% Top-1 and 45.0% Top-5 high-quality prediction success, together with an acceptable-or-better success rate of 72.5% for Top-1 predictions, on the evaluated 160-interface FoldBench antibody–antigen subset. By combining broad target coverage, repeatable design improvement, and transfer across binder formats, MMDesign provides a reusable discovery platform for diverse molecular engineering programs with a small experimental candidate set.
Files
MMDesign Technical Report-2026Oct11.pdf
Additional details
Related works
- Obsoletes
- Report: 10.5281/zenodo.20526589 (DOI)
Software
- Repository URL
- https://mos.moleculemind.com/
- Development Status
- Active