Microsoft Research has published RetroChimera, a retrosynthesis prediction framework presented in a Nature paper, and open-sourced its implementation and weights under the MIT license. For chemists working on medicines and advanced materials, the model is intended to improve the ranking of reaction steps used to plan routes from a target molecule back to purchasable building blocks.
Retrosynthesis works backward by repeatedly dividing a desired molecule into simpler precursors. Because many disconnections may be possible at every step, searches can branch into numerous incomplete routes. Automated systems must recover rare but strategically useful reactions, work beyond their training distribution and recommend steps that practicing chemists consider credible; one incorrect step can invalidate an otherwise promising multistep route.
Complementary models and learned voting
RetroChimera combines two models with different strengths. R-SMILES 2 is a Transformer-based de novo model that predicts precursor molecules directly from the target. This unconstrained approach can learn reaction patterns from data and performs particularly well on transformations involving large changes, but Microsoft notes that it can also hallucinate.
NeuralLoc instead represents the target molecule and reaction templates as graphs. The graph neural network selects a template and identifies where it should be applied. Grounding predictions in patterns extracted from training data generally makes its output more reliable, according to Microsoft, especially for low-precedence reactions and transformations involving localized changes. The trade-off is that NeuralLoc is constrained when a required reaction is absent from its template library.
The framework aggregates the models’ ranked proposals through a learned voting system. Each component assigns a rank-dependent vote to every predicted reactant set, and the votes are added when both components propose the same reaction. Learning how much weight to place on each model at different ranks lets the ensemble exploit their distinct inductive biases. Microsoft reports that RetroChimera approximately matches the better-performing component across reaction classes instead of inheriting the weaknesses of either approach uniformly.
What the evaluations found
In blind assessments, PhD-level chemists preferred RetroChimera’s proposed disconnections for complex molecules over outputs from its constituent models, more established approaches and reactions recorded in the test set. The source also reports strong performance across common and rare reaction classes, while the broader validation covered recall of rare reaction types, zero-shot transfer and fine-tuning on proprietary datasets.
The multistep evaluation involved ten challenging targets. Experts accepted RetroChimera routes for nine targets, compared with five for the de novo model, four for the editing model and two for NeuralSym, the baseline. The accompanying results describe a 90% acceptance rate for RetroChimera’s routes, versus 20% to 50% for routes produced by its sub-models or the baseline. These figures measure expert acceptance in the supplied evaluation rather than successful execution of every route in a laboratory.
Microsoft argues that better route proposals could help researchers examine more candidate molecules and shorten design-make-test cycles in drug discovery and smart-material development. It also anticipates that synthesis planning could eventually be paired with laboratory automation in closed-loop systems. Those are prospective applications: the evidence supplied here supports prediction quality and alignment with expert judgment in the reported tests, not a measured reduction in laboratory time or cost across arbitrary molecules. Microsoft is asking the chemistry community to test the framework further and identify its strengths and shortcomings.
RetroChimera is available through GitHub, with checkpoint-access instructions in the repository, and through Microsoft Foundry. The full study appears in Nature.
Source: Microsoft Research
Definition. RetroChimera is a retrosynthesis prediction framework that uses learned voting to combine the de novo R-SMILES 2 model with the template-based NeuralLoc model.
| System | Expert-accepted routes |
|---|---|
| RetroChimera | 9 of 10 targets |
| De novo model | 5 of 10 targets |
| Editing model | 4 of 10 targets |
| NeuralSym baseline | 2 of 10 targets |
Key takeaways
- RetroChimera’s implementation and weights were released under the MIT license.
- The framework combines R-SMILES 2’s unconstrained predictions with NeuralLoc’s template-grounded proposals.
- A learned voting system aggregates and weights ranked proposals from both models.
- PhD-level chemists preferred RetroChimera disconnections over outputs from constituent models, established approaches and test-set reactions in blind assessments.
- Experts accepted routes for nine of ten multistep targets, compared with two to five targets for the evaluated alternatives.
- Claims about shorter laboratory timelines and lower costs remain prospective rather than demonstrated.
FAQ
What is RetroChimera used for?
It predicts retrosynthetic routes by working backward from a target molecule toward simpler, purchasable building blocks.
How does RetroChimera combine its models?
It assigns rank-dependent votes to proposals from R-SMILES 2 and NeuralLoc, adding their votes when both models predict the same reactant set.
How did RetroChimera perform in the multistep evaluation?
Experts accepted its routes for nine of ten challenging targets, corresponding to a reported 90% acceptance rate.
Does the evaluation prove that RetroChimera speeds up laboratory synthesis?
No. The reported results support prediction quality and alignment with expert judgment, not a measured reduction in laboratory time or cost.
Where is RetroChimera available?
It is available through GitHub, with checkpoint-access instructions in the repository, and through Microsoft Foundry.
