Microsoft Research Open-Sources RetroChimera for Chemical Synthesis Planning

Microsoft Research has open-sourced RetroChimera, an AI system for working backward from a target molecule to propose routes for making it. The Nature-published work joins two models with different limits, then learns how to rank their suggestions into one synthesis-planning output.

Retrosynthesis begins with a desired molecule and repeatedly breaks it into simpler chemical precursors, aiming to reach building blocks that can be bought. The difficulty is not just proposing one plausible reaction: a useful route requires choosing a sequence of disconnections from a very large set of possibilities.

RetroChimera combines R-SMILES 2, a Transformer model that predicts precursor molecules directly, with NeuralLoc, a graph neural network that selects reaction templates and where to apply them. Microsoft says the first approach can learn patterns freely but can generate hallucinated outputs; the second is grounded in templates from training data but is limited by what its template library covers.

Rather than selecting one method, the system has both models rank possible reactant sets. It combines those rankings with learned, rank-dependent votes, adding weight when both models propose the same reaction. Microsoft says this lets the ensemble approximately match the stronger sub-model across reaction classes, including common and rare ones.

Microsoft also reports that PhD-level chemists preferred RetroChimera’s individual reaction predictions over preceding models and recorded literature reactions in blind tests. Those are promising results, but the full-route comparison covers ten challenging targets, so researchers can now test how the released system performs on the molecules and reaction types relevant to their own work.

RetroChimera’s implementation and weights are available on GitHub under the MIT license, and the system is also accessible through Microsoft Foundry. The release gives researchers access to the same ensemble design behind the reported evaluations.