Chinese AI drug discovery company Molecule Hearts has released the latest technical report for MMDesign, its de novo protein design platform, according to QbitAI: across 19 design projects spanning 18 targets and two molecular modalities — nanobodies (VHH) and macrocyclic peptides — 18 projects yielded bindable molecules, a 94.7% project success rate.
A second set of numbers matters just as much: of 453 candidates tested at the bench, 139 confirmed binding, an overall hit rate of 30.7%. The two gauges measure different things — 94.7% asks whether a project finds at least one binder; 30.7% asks how many of the molecules sent to the lab actually bind. In de novo design, both figures currently sit at the top tier.
The efficiency comes from a compressed workflow. MMDesign fixes the process into generate-screen-validate: it first generates tens of thousands of candidates in computation, then applies layered screening on structural reliability, sequence naturalness and statistical physics, compressing each project's candidate set to 9-50 molecules with a median of just 21 — fewer molecules enter wet-lab validation, and per-project cost falls accordingly.
The route carries a second advantage: it needs no known binder as a reference. Sequences are generated directly from functional requirements, so even targets without precedent produce answers — the essential difference between de novo design and re-engineering known molecules. QbitAI's report also notes its precision on hard antibody design reaches twice that of AlphaFold 3.
Three iterations in six months suggest MMDesign is turning high hit rates from a one-off result into a repeatable capability. The signal matters more than any single figure: AI protein design has moved from proof-of-concept papers to systematic validation across targets and modalities.