The Pulse
Microsoft Open-Sources RetroChimera for Molecule Synthesis Planning
Microsoft Research has released RetroChimera, an MIT-licensed AI model that predicts retrosynthetic routes for small molecules. The system combines two complementary models and outperformed several alternatives in expert evaluations involvi
AI.info Team ·
Microsoft Research is releasing the code and model weights for RetroChimera as the company argues that AI can improve a task chemists still handle through laborious, expert-led planning: working backward from a desired molecule to the ingredients and reactions needed to make it.
The release comes alongside a Nature paper describing the model and its evaluation. Microsoft presents RetroChimera as a response to a persistent weakness in computer-assisted synthesis planning: systems can perform well on common reactions yet miss rare transformations, generate chemically implausible suggestions, or produce routes that fail when several steps are chained together.
That claim comes with a qualification from Microsoft’s own repository. The company warns that RetroChimera can hallucinate, particularly outside its training distribution, and says all predictions must be assessed and independently verified by chemistry experts. The tension between stronger benchmark results and the need for expert supervision defines the release.
Two models cover different chemical blind spots
RetroChimera predicts retrosynthetic steps. Given a target molecule, it works backward by proposing precursor molecules and reactions that could produce them, repeating the process until a route reaches simpler or commercially available building blocks.
The system combines two sub-models with different designs. R-SMILES 2 is a Transformer-based model that generates precursor molecules directly from the target structure. Its freedom to learn reaction patterns from data helps it handle transformations involving substantial changes to a molecule, but that same freedom can produce invalid outputs.
NeuralLoc takes a more constrained approach. The graph neural network represents the target molecule and reaction templates as graphs, then predicts which template should be applied and where. Template-based predictions can be more reliable when the relevant chemistry appears in the training data, although the method is less flexible when a reaction falls outside its template library.
RetroChimera uses learned ranking to combine their proposals. Each model assigns rank-dependent scores to candidate reactant sets, with additional support when both models propose the same reaction. Microsoft says the ensemble can match the stronger component across different reaction classes while reducing the weaknesses of either model used alone.
Experts accepted nine of ten complete routes
Microsoft’s evaluation focused on a difficult part of synthesis planning: a single incorrect reaction can invalidate an entire multistep route. In an expert assessment of 10 challenging targets, chemists accepted complete routes from RetroChimera for nine targets. The corresponding totals were five for the de novo model, four for an editing model and two for the NeuralSym baseline.
Individual reaction predictions also performed well in blind comparisons. Microsoft says Ph.D.-level chemists preferred RetroChimera’s suggestions over outputs from its component models, established approaches and reactions drawn from the test set. A separate Microsoft account of the work says nine organic chemists from Microsoft and major pharmaceutical companies preferred RetroChimera’s top proposal to previously documented routes about 64% of the time.
Those results measure whether experts consider a proposed reaction or route acceptable. They do not show that the model has independently synthesized every proposed compound, and they do not remove the need for laboratory testing. Microsoft’s repository specifically advises users to request no more than five to 10 reactions per input unless they apply stringent filtering.
Proprietary data is the harder test
Public reaction datasets are useful for benchmarks but do not represent every company’s chemistry. Pharmaceutical teams often work with internal electronic laboratory records and reaction collections that differ from public training data in the molecules, reaction classes and documentation practices they contain.
The Nature study reports zero-shot transfer to an internal dataset from a major pharmaceutical company, while Microsoft’s account says the model was adapted to GSK’s proprietary data. The company frames that result as evidence that RetroChimera can move beyond standard academic benchmarks and support settings where the available chemistry differs from the data used to build the original model.
Microsoft does not present the release as a finished autonomous chemist. The repository says the main RetroChimera checkpoint was trained on Pistachio reaction data and that the training data available to the model runs through 2023. Microsoft also supplies weaker checkpoints trained on USPTO-50K and USPTO-FULL for benchmarking, along with a forward-model checkpoint trained on Pistachio.
MIT license, but expert review stays mandatory
The implementation is available in Microsoft’s public GitHub repository under the MIT license, and Microsoft says the model can also be accessed through Microsoft Foundry. The repository includes installation instructions, inference examples and guidance for using the released checkpoints.
Microsoft’s documentation lists several limitations. Predictions may be less accurate for specialized academic chemistry and natural-product chemistry, while noise in the Pistachio dataset means outputs should be checked against relevant literature. The company recommends pairing the model with reaction-feasibility tools, using consensus modes and fine-tuning on domain-specific laboratory or literature data where appropriate.
That makes RetroChimera a research system for prioritizing and testing synthesis ideas, not a replacement for chemical judgment. Its immediate contribution is more concrete: researchers can now inspect, run and adapt the same ensemble framework described in the Nature study rather than relying only on published performance claims.
RetroChimera’s code and weights are available through GitHub, with the model’s research findings described by Microsoft Research.