Microsoft Research has released RetroChimera, an AI system that helps chemists work out how to make a target molecule from simpler starting materials. It combines two models that approach this planning task differently, improving predictions for reactions that appear infrequently in training data. Following publication in Nature, the code and trained models are available for outside researchers to test.
The task is called retrosynthesis: starting with the molecule a chemist wants, a planner works backward through possible reactions until it reaches materials that can be obtained. A useful plan may need several steps. If one proposed reaction is implausible, the entire route can fail, so each suggestion has to be credible even when the relevant chemistry is uncommon.
One part of RetroChimera, R-SMILES 2, generates possible starting materials from a written representation of the target molecule's structure. It uses a Transformer, a neural network that learns relationships within sequences. This gives the model room to propose reactions beyond a fixed library of rules, but also leaves it able to invent chemically implausible suggestions.
The second model, NeuralLoc, supplies a more constrained approach. It treats molecules as networks of atoms and bonds, chooses a reaction pattern learned from previous examples, and identifies where that pattern fits the target. This can help with reactions supported by only a few examples, although its library limits what it can propose. RetroChimera then combines the two lists, learning how much weight to give each model's suggestions at different positions in the ranking.
The researchers tested the combination on patent-derived reaction datasets, a larger curated collection and pharmaceutical-company data. In tests grouped by how often a reaction pattern appeared during training, the two models had different strengths; combining them improved coverage of uncommon chemistry. These tests measure whether the suggested starting materials match a recorded reaction. A different answer may also be useful, which is why the team asked chemists to assess the proposals.
In a small review of complete routes for ten challenging molecules, experts accepted nine RetroChimera routes, compared with two to five from the other tested models. Microsoft reports this as expert assessment of the plans, not confirmation that all the proposed routes succeeded in laboratory experiments.
The public release lets other researchers examine those strengths and limits using an MIT-licensed implementation and downloadable trained models, also called checkpoints. The strongest released checkpoint learned from the Pistachio reaction collection using data available through 2023. The repository warns that unfamiliar chemistry and suggestions lower in the output list are more likely to produce errors.
RetroChimera's output therefore remains a starting point for a chemist's planning. The reported task proposes starting molecules without specifying reagents, solvents or reaction conditions. Choosing those details, checking the literature and verifying the chemistry experimentally still separate an accepted plan from a working synthesis.
Sources
- Nature: Chemist-aligned retrosynthesis by ensembling diverse inductive bias models
- Microsoft Research: Improving synthesis prediction of small molecules at scale with RetroChimera
- GitHub: microsoft/retrochimera code and model access instructions
- Representative laboratory glassware photograph by Tweenk, CC BY 3.0. The image does not show RetroChimera or a reaction planned by it.