MoleculeMind claims 94.7% success rate in AI protein design projects

Plenty of players can already generate thousands of protein sequences. MoleculeMind (分子之心 in Chinese) is proposing to measure something different: how many molecules need to be sent off for experimental validation, and how often a project actually ends with a working candidate. In a new technical report on its AI design platform MMDesign, the company cites a 94.7% project success rate across 18 targets. All six GPCR targets in the set were successfully hit.

Why the key metric is attrition, not generation

In AI-driven biologics development, the bottleneck has long since shifted. A model can churn out as many variants as you like, but every experimental cycle costs money and weeks of work: synthesis, expression, affinity measurement. If only a handful out of a thousand submitted molecules turn out to work, the project economics collapse no matter how impressive the generation step looks.

That's why MoleculeMind is shifting the emphasis to the "lab entry point": the fewer candidates needed to be tested for a success, the cheaper each project becomes. The 94.7% figure is measured by project, not by individual molecule — and that reflects exactly this logic. The platform is being evaluated as a pipeline, not a generator.

Six GPCRs as a stress test

The most striking part of the report concerns GPCRs — G-protein-coupled receptors, which account for a significant share of approved drugs. For biologics, they remain a difficult target. The receptor sits in the membrane, its extracellular regions are small and mobile, and the purified protein is often unstable. Antibodies against such targets tend to come out inconsistently.

That's why a set of six GPCR targets, each of which the platform successfully hit, looks far more convincing than success on "easy" soluble proteins. It marks a shift from one-off demonstrations to systematic validation — as the report's own title suggests.

What exactly was tested

Key parameters from the report:

• platform: MMDesign, an AI system for de novo molecule design;
• developer: MoleculeMind (分子之心);
• scope: 18 targets;
• modalities: nanobodies (VHH) and macrocycles;
• project success rate: 94.7%;
• GPCR targets: all six were hit;
• report publication date: October 27, 2026.

The combination of two modalities is also telling. VHHs are single-domain antibodies — more compact than classical immunoglobulins and better at reaching hard-to-access regions of a target. Macrocycles sit somewhere between small molecules and biologics. A single platform capable of handling both protein-based and cyclic structures reduces dependence on any one format.

How to read claims like this

The report was published by the company itself, and independent verification is not yet available. That's why the details determining the real value of that 94.7% matter so much to the market:

• how "project success" is defined: binding at the level of a few candidates, functional activity, or affinity within a target range;
• how many molecules had to be synthesized and tested per target;
• how novel the targets are relative to the training data and known antibodies;
• whether there are results beyond "hits": selectivity, stability, suitability for further development.

For pharma companies and investors, this amounts to a practical checklist. If the answers to these questions hold up, MMDesign would become a strong argument that AI-driven biologics design is moving from demonstrations to a predictable process. For teams building similar platforms themselves, the benchmark is now clear: compare the share of successful projects and the number of molecules that actually went into experiments, rather than the sheer volume of generation.