A new tool makes AI models better at figuring out how to build molecules, without touching the models themselves.
Retrosynthesis prediction is the process of working backward from a target molecule to figure out what starting materials and reactions could produce it. It matters most in drug discovery, where chemists need to know if a promising compound can actually be made. Researchers built RetroMPA, a module that sits on top of existing retrosynthesis AI models and re-ranks their predictions using known chemical properties, rather than replacing them. Tested across eight different models on a standard chemistry dataset (USPTO-50K), it lifted top-1 accuracy by an average of 5.50%, and by about 2.03% on a much larger dataset. The team also ran real lab syntheses and says they confirmed new, previously unreported substrate combinations for three well-known reaction types.
The interesting part is the framing: this isn't a bigger or fancier model, it's a knowledge-based patch for models that already exist. Most gains in this field have come from throwing more data and parameters at the problem. RetroMPA instead argues that raw pattern-matching on reaction data misses chemistry that scientists already understand, and that injecting that knowledge back in is cheaper than retraining. If the wet-lab results hold up under more scrutiny, that is a meaningful claim, not just a benchmark bump.
Code is open-source, so anyone running one of those eight base models can bolt this on and check the accuracy claims themselves.