Microsoft Research reported the publication of its retrosynthesis model RetroChimera in Nature. RetroChimera is a framework for retrosynthesis prediction built around two models. Retrosynthetic planning starts from a target molecule and considers possible precursors or disconnections that could lead to it. The framework approaches that task by combining two different modeling methods and then bringing their ranked suggestions together.

Two models, different routes to suggestions

The first component, R-SMILES 2, is a Transformer-based de-novo model. It predicts precursor molecules directly from the input molecule. Its described role is direct prediction: it supplies one route from a target molecule to possible precursors without relying on the template-selection procedure described for the other component.

The second component, NeuralLoc, is a graph neural network. It encodes target molecules and reaction templates as graphs, selects templates, and predicts where to apply them. Its process therefore makes reaction-template choice and placement part of the prediction. The two components offer different ways to propose retrosynthetic outcomes, and RetroChimera uses ranked predictions from both rather than presenting only one model's list.

Combining ranked predictions

RetroChimera combines the ranked predictions of its two sub-models through a learned ensembling strategy. The strategy uses rank-dependent votes, so a prediction's position in a model's ranking matters when the results are combined. When both models propose the same reaction, their votes are added. Agreement between the components therefore contributes to the combined ranking, alongside the ranks assigned by the individual models.

This is the central design described in the article: one component predicts precursors directly, while the other selects and locates reaction templates; a learned procedure combines their ranked outputs. The combination is a way of organizing predictions from the two models. The article's description does not make the ranking a guarantee that a proposed route has been experimentally carried out or that every suggestion is feasible. A prediction is a proposal for consideration, and the reported expert assessment concerns preferences among suggestions.

Reported performance and expert assessment

The article says RetroChimera performs strongly across common and rare reaction classes and produces predictions that better align with chemists' judgment. It also reports blind tests in which expert chemists preferred RetroChimera's suggestions for disconnections of complex molecules over suggestions from its constituent sub-models, established approaches, and the test set itself. This comparison reports what the participating experts preferred in those tests; it should be read within that scope.

The distinction matters for understanding what the result says. The system predicts retrosynthetic suggestions, and the blind-test account describes expert preference among those suggestions. The article does not state that each proposed disconnection was experimentally validated. Accordingly, the reported preference does not by itself establish that every ranked proposal can be carried out in a laboratory. The result described by Microsoft Research is that the combined model's suggestions were preferred in the stated comparison, alongside its reported performance across common and rare reaction classes.

Availability and context

Microsoft Research said it open-sourced RetroChimera's implementation and weights. The article states that the model is available on GitHub under an MIT license and accessible through Microsoft Foundry. These details identify the availability channels reported by Microsoft Research; they do not add a separate performance claim. The work is presented as research on predicting synthesis of small molecules at scale, with the framework's two-model design at its center.

For readers, the practical point is the division of labor within the model. R-SMILES 2 maps an input molecule directly to predicted precursor molecules. NeuralLoc represents targets and templates as graphs, chooses templates, and predicts their application sites. RetroChimera then uses learned rank-dependent voting to combine the two ranked outputs, adding votes when both models propose the same reaction. The reported blind-test preference provides an expert comparison of these suggestions, while the publication and availability statements describe how Microsoft Research presented and released the work.

Together, these elements explain the article's account of RetroChimera: a retrosynthesis-prediction framework, a combination of direct generation and template-based prediction, a learned ranking strategy, and reported expert preferences for complex-molecule disconnections. The claims here follow the Microsoft Research article, including its statements about publication, reported performance, and access to the implementation and weights.