Chemistry World covered the work under the headline "Microsoft's split-personality retrosynthetic AI is a fast learner." The nickname fits, but the useful questions are how the model works, what the results actually show, and where its limits are.
What retrosynthesis is
Chemists usually think about a synthesis backwards. Retrosynthesis is a technique that starts with a desired target molecule and works backward, step by step, to break it down into simpler, commercially available building blocks. Microsoft compares it to chess and Go, but notes that its decision space is vastly larger and more complex.
That complexity has held automation back. According to Microsoft, for decades, this immense combinatorial complexity led researchers to believe that retrosynthesis could not be automated reliably. In Microsoft's research summary of the paper, earlier models are described as still struggling with reactions that are rare but important to a synthetic strategy, and with hallucinated, incorrect outputs.
Section summary: Retrosynthesis takes a target molecule and works back to a route made of buyable starting materials. Earlier AI tools had trouble with rare reactions and invented answers.
Two models with different strengths
The name refers to the mythological chimera, a creature made of parts from different animals. In Microsoft's words, RetroChimera is built around two models:
- R-SMILES 2. A Transformer-based de-novo model, predicts precursor molecules directly from the input molecule. SMILES (Simplified Molecular Input Line Entry System) writes a chemical structure as a string of letters and symbols. That lets R-SMILES 2 treat chemistry much as a language model treats text. Microsoft also points out the downside: its unconstrained generation can also make it prone to hallucination.
- NeuralLoc. This is a graph neural network. Microsoft says it encodes the target molecule and a set of reaction templates as graphs, picks a template, and predicts where on the molecule to apply it. Because its suggestions come from reaction patterns found in the training data, they tend to be more reliable. The cost is that it has trouble with reactions its template library doesn't cover.
The trade-off between the two is the point. Microsoft reports that R-SMILES 2 performs particularly well on reactions that involve large changes over the course of the reaction, while NeuralLoc excels in reactions of low precedence and those involving more localized changes.
The model doesn't just alternate between their answers. RetroChimera combines the ranked predictions of both sub-models using a learned ensembling strategy. Microsoft says each model casts a learned vote whose weight depends on rank. When both models propose the same set of reactants, their votes add up. Over time the system learns how much to trust each model at each position in its ranked list.
A simple way to picture it: one reviewer is very inventive but sometimes makes things up, the other follows established precedent and misses anything new. The ensemble learns when to listen to which.
Section summary: RetroChimera combines a sequence-based generator (R-SMILES 2) with a template-based graph network (NeuralLoc). A learned voting system ranks their combined suggestions.
What the evaluations show
The research paper, "Chemist-aligned retrosynthesis by ensembling diverse inductive bias models," appeared in Nature. Microsoft's research summary says experiments ran across several orders of magnitude in data scale and splitting strategy. According to that summary, RetroChimera beat the major existing models by a wide margin, held up on data outside its training set, and could learn from very few examples per reaction class.
The results come from several separate tests, and they shouldn't be mixed together:
| Evaluation | Reported result | Source |
|---|---|---|
| Single-step preference test | Nine PhD organic chemists preferred RetroChimera's top suggestion over the documented reaction about 64% of the time | Microsoft Source |
| Whole-route test, 10 hard targets | Complete routes accepted for 9 targets, vs. 2–5 for the comparison models | Microsoft Research / Source |
| Novartis internal data | Direct retrosynthesis on 10,444 company reactions | Chemistry World |
| GSK internal data | More than 650,000 reactions used for training, validation, testing and fine-tuning experiments | Chemistry World |
| Hallucinations | Fewer than half as many as without cross-checking between the submodels | Chemistry World |
On the full-route test, Microsoft says RetroChimera succeeded on nine of the ten targets. It compares that with five for the de novo model, four for the editing model and two for NeuralSym, an established baseline. Chemistry World describes the pharmaceutical work this way: Team members from Novartis in Basel, Switzerland, enabled the researchers to use Retrochimera directly to do retrosyntheses on 10,444 reactions from the drug company's internal database, and GSK in Stevenage provided over 650,000 reactions from its database for training, validation and testing.
Chemistry World also reports that the combined model beat both submodels throughout, especially on rare transformations, and that the cross-checking between submodels decreases the number of hallucinations, or strange suggestions, by more than half. Microsoft's own summaries explain why combining the models should reduce hallucinations, but they don't give that figure. Treat "more than half" as Chemistry World's reported result from a specific evaluation, not a guarantee for every input.
Philippe Schwaller of EPFL, who was not involved, told Chemistry World he considers RetroChimera arguably the strongest single-step retrosynthesis model so far. He also argued that the field's usual scoring method, checking whether a model reproduces the reaction recorded in a database, is a flawed yardstick because a molecule usually has more than one valid route. He would like blinded expert evaluations to become standard practice.
A necessary caveat: chemists preferring a suggestion is not the same as the reaction working in a flask. These are expert ratings of predicted reactions. Microsoft itself says the next step is to evaluate the model in real-world discovery settings, moving beyond benchmark testing.
Section summary: The model did well in blind chemist reviews and on industry data. These are research evaluations, not lab-confirmed syntheses.
How to try it: GitHub, Python and Foundry
Developers can use the model in two ways.
Microsoft Foundry. The Foundry catalog entry is listed as Version 2, credited to Microsoft Research AI for Science, and Model type: Chemical Reaction Prediction Model - Language(s): Python - License: MIT. The input is a product molecule written as a SMILES string. Each reaction is represented as a group of ingredients (reactant molecules), with those molecules again represented each by a SMILES string.
Running it locally. The GitHub repository describes this setup:
- Create the conda environment from the repository's
environment.ymlfile:conda env create -f environment.yml - Activate it with
conda activate retrochimera - Install the package with
pip install retrochimera - Download a model checkpoint. Inference won't run without one, and the repository explains how to get them.
- In Python, load
RetroChimeraModeland point it at the checkpoint directory. Pass in a target molecule as asyntheseusMoleculebuilt from a SMILES string, then request a set number of results. The repository's example asks for three, and each prediction comes with a probability.
Checkpoints and dependencies:
- The main (and most powerful) checkpoint we release is trained on Pistachio. For benchmarking, there are also weaker checkpoints trained on USPTO-50K and USPTO-FULL, plus a forward-prediction model that predicts products from reactants.
- To run the USPTO-50K checkpoint, install the extra dependency group:
pip install retrochimera[graphium]. - To reproduce the paper's USPTO results exactly, use the inference hyperparameters listed in the paper's Extended Data Tables 3 and 4. The repository notes that by default, these parameters are set to values optimized for the Pistachio checkpoint.
The Foundry Labs page suggests pairing it with Microsoft's other science models, saying the public checkpoints pair naturally with property predictors and generators such as TamGen and MatterGen.
Section summary: Install from GitHub with conda and pip, or use the Foundry catalog. Either way you need a checkpoint, and the Pistachio one is the strongest.
Limits to know before using it
The repository is unusually upfront about risk. RetroChimera 1 is released for research and experimentation, and its documented warnings include:
- Expert checks are required. Predictions must be risk-assessed and independently verified by chemistry experts before any real-world use.
- Lower-ranked results are less reliable. Suggestions further down the list are more likely to be hallucinations. The project recommends asking for no more than five to ten reactions per input unless you add strict filtering.
- Training data stops in 2023. The Pistachio checkpoint doesn't include newer chemistry.
- Some chemistry is out of scope. The repository warns that performance may be lower for chemistry that differs substantially from the Pistachio training distribution, including specialized academic and natural-product chemistry.
- Recommended practice: pair it with a reaction-feasibility model, use consensus mode, and fine-tune on relevant electronic lab notebook data or literature reactions for specialist work.
For teams running this in a regulated R&D environment, those warnings map directly onto a deployment plan. You need a human approval step, a confidence threshold, and a fine-tuning pipeline whose own security and data-governance requirements cover proprietary reaction data.
What's next
Marwin Segler, who leads the work at Microsoft Research AI for Science in Cambridge, told Chemistry World that fine-tuning on internal data improves results further, which matters for pharmaceutical and chemical companies. He added that because the system is modular, new submodels can be added later. He also said Microsoft has continued improving its internal models with collaborators since the paper was submitted. Other projects he mentioned are Qfang, which predicts reaction conditions, and an AI agent called Clio that uses RetroChimera.
One timeline note: the work isn't brand new. The paper appeared as an arXiv preprint before its Nature publication, and that preprint already said RetroChimera is available through Azure Foundry. Segler says it is still competitive in independent benchmarks. Others are building on it too: one recent arXiv framework, RETROSPECT, describes its proposal model as something that can also be inserted into ensemble systems such as RetroChimera for future work.
Bottom line: RetroChimera is a good example of combining models instead of just building a bigger one. The pairing addresses specific weaknesses, it holds up on real pharmaceutical data, and it's openly licensed. It suggests possible routes; whether a reaction actually works is still something chemists have to check.