Microsoft Research said on 21 September that its retrosynthesis model RetroChimera has been published in Nature, and released the implementation and weights under an MIT licence. Retrosynthesis works backwards from a target molecule through successive disconnections until it reaches purchasable starting materials. It is the planning step that decides whether a candidate drug or material is practical to make at all.
RetroChimera is an ensemble of two models with opposite failure modes. R-SMILES 2, a transformer, generates precursors directly and handles reactions involving large structural change, but its unconstrained generation can hallucinate. NeuralLoc, a graph neural network, picks from a library of reaction templates and is more reliable on localised or rarely precedented reactions, but is limited to what that library covers. A learned ranker weighs the two and adds their votes where they agree.
In blind tests, expert chemists preferred RetroChimera’s individual disconnections to those of either sub-model, to established approaches, and to the reactions actually recorded in the literature. On complete routes it succeeded on nine of ten challenging targets, against five for the de novo model, four for the editing model and two for a strong baseline.
Ten targets is a small sample, and a chemist’s preference is a judgement rather than a laboratory result. That the weights are public is what makes the claim checkable.