Chemistry & ADMET · Microsoft Research AI for Science
RetroChimera
Proposes ranked one-step retrosynthetic disconnections for target molecules using an ensemble of two distinct models.
What RetroChimera is
RetroChimera is a one-step retrosynthesis system developed by Microsoft Research AI for Science. Retrosynthesis is the practice of planning chemical syntheses by recursively deconstructing target molecules into simpler, readily available starting materials. The system tackles the challenge of generating practical, high-quality reaction disconnections for synthetic chemistry.
The model couples two submodels with contrasting inductive biases: one model generates Simplified Molecular Input Line Entry System (SMILES) strings of the reactants directly, whereas the second localises reaction sites and applies discrete reaction templates. Their individual predictions are combined into a single prioritised output list, allowing agreement between the submodels to act as a confidence metric.
In blind evaluations, industrial organic chemists preferred the candidate disconnections suggested by RetroChimera over the historical literature reactions present in its training corpus. Additionally, the system demonstrates the ability to acquire new reaction classes from very few training examples.
How it works
RetroChimera combines two different algorithmic approaches to retrosynthesis. One component is a template-free model that directly outputs reactant SMILES strings. The other component is a template-based model that identifies the reactive centre of the substrate and applies matching reaction transformation rules. Predictions from both components are merged and ranked into a unified disconnection list.
The software supports a consensus mode, which restricts results strictly to disconnections proposed by both submodels, offering higher precision at the expense of candidate diversity. The primary checkpoint was trained on the Pistachio reaction database through 2023, while additional benchmarking checkpoints are provided for the USPTO-50K and USPTO-FULL datasets. A companion forward-reaction prediction checkpoint based on the SMILES-generation architecture is also available.
Key facts
- Version
- 1.3
- Domain
- Chemistry
- Tasks
- Retrosynthesis, reaction prediction
- Modalities
- Small molecule
- Input formats
- Target molecule SMILES, ranking mode, number of results
- Output formats
- Table of ranked reactant SMILES, confidence scores, submodel agreement counts
- Primary training corpus
- Pistachio (reactions through 2023)
- Benchmark datasets
- USPTO-50K, USPTO-FULL
- Required GPU memory
- 24 GB
What you can use it for
- One-step retrosynthesis planning
- Proposing ranked precursor sets and disconnections for target small molecules during synthetic route design.
- High-precision candidate filtering
- Applying consensus mode to filter out low-confidence reactions, retaining only disconnections agreed upon by both submodels.
- Low-data reaction class adaptation
- Extending retrosynthetic predictions to novel or rare reaction classes represented by very few training examples.
- Forward reaction viability checking
- Assessing whether proposed reactant combinations produce the intended target using the matched forward-prediction checkpoint.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
RetroChimera
Strengths and limits
Good at
- Industrial organic chemists favoured RetroChimera suggestions over historical training literature in blind evaluations.
- The ensemble architecture fuses template-free SMILES generation and template-based logic to provide a clear agreement signal.
- The model learns new reaction classes effectively from very few examples per class.
- Consensus mode provides a simple mechanism to boost precision by discarding disconnections that lack mutual submodel support.
Watch out for
- Lower-ranked suggestions are increasingly prone to hallucination, making expert chemical review or reaction-feasibility filtering essential.
- Training data for the primary checkpoint ends in 2023, omitting newly reported chemical reactions.
- Model performance may drop on substrates that diverge significantly from the Pistachio corpus, such as natural products or specialized academic chemistries.
Inputs and outputs
You provide
Target molecule (SMILES)
One product, not a reaction and not a mixture. Up to about 200 heavy atoms.
How to rankoptional
Disconnectionsoptional
Capped at ten: the authors report that suggestions past that point are increasingly likely to be hallucinations.
You get back
Ranked reactant sets with the ensemble's confidence and how many submodels proposed each. Every SMILES is checked for whether it parses.
Licence
Commercial use permitted · MIT code / MIT weights
Verified 2026-10-29. Code: MIT, github.com/microsoft/retrochimera/blob/main/LICENSE, and the `retrochimera` PyPI package carries the same text. Weights: the three figshare records holding the checkpoints each declare MIT — figshare.com/articles/software/RetroChimera_Pistachio_/30591107 (the one that runs here), plus the USPTO-FULL and USPTO-50K records. Neither the code nor the weights are gated. Note on provenance rather than licence: this checkpoint was trained on Pistachio, a commercially licensed reaction database, and Microsoft released the resulting weights under MIT; the weights carry no clause inheriting Pistachio's terms.
Questions
Can I use RetroChimera commercially?
Commercial use permitted. Verified 2026-10-29. Code: MIT, github.com/microsoft/retrochimera/blob/main/LICENSE, and the `retrochimera` PyPI package carries the same text. Weights: the three figshare records holding the checkpoints each declare MIT — figshare.com/articles/software/RetroChimera_Pistachio_/30591107 (the one that runs here), plus the USPTO-FULL and USPTO-50K records. Neither the code nor the weights are gated. Note on provenance rather than licence: this checkpoint was trained on Pistachio, a commercially licensed reaction database, and Microsoft released the resulting weights under MIT; the weights carry no clause inheriting Pistachio's terms.
How do I run RetroChimera without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 25s, and a run costs about 14 credits. Visitors without an account get 1,000 credits a month.
What is the difference between the two submodels?
One submodel is template-free and writes the SMILES strings of precursors directly, while the other is template-based, identifying the reactive bond before applying established reaction templates.
How many disconnections should be generated per molecule?
It is recommended to request no more than 5 to 10 reactions per target molecule, because lower-ranked candidates are increasingly prone to hallucinations unless filtered.
What does consensus mode do?
Consensus mode retains only those candidate disconnections that were independently proposed by both the template-free and template-based submodels, increasing precision.
Can RetroChimera predict forward reactions?
A forward model checkpoint is released alongside the retrosynthesis model, sharing the architecture of the SMILES-generating submodel and trained on the Pistachio corpus.
How to cite it
Maziarz, Krzysztof, Guoqing Liu, Hubert Misztela, Austin Tripp, Junren Li, Aleksei Kornev, Piotr Gaiński, Holger Hoefling, Mike Fortunato, Rishi Gupta, and Marwin Segler. 'Chemist-aligned retrosynthesis by ensembling diverse inductive bias models.' arXiv preprint arXiv:2412.05269 (2025).
Written from RetroChimera's own documentation (code, paper). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.