Microsoft's RetroChimera Plans Drug Synthesis Routes and Wins 9 of 10 Expert Reviews
Microsoft Research's RetroChimera combines two neural networks to plan chemical syntheses that expert chemists prefer over published literature reactions.
- Microsoft Research published RetroChimera in Nature, a retrosynthesis prediction model.
- Ensembles a SMILES Transformer (R-SMILES 2) with a graph neural network (NeuralLoc) via learned voting.
- PhD chemists preferred its disconnections over sub-models, baselines, and literature reactions in blind tests.
- Solved 9 of 10 hard multistep targets versus 2-5 for competing models.
- Supports zero-shot transfer and fine-tuning on proprietary reaction datasets.
- Open-sourced on GitHub under MIT license and available on Microsoft Foundry.
Finding a promising drug or advanced material leaves researchers with an expensive practical problem: planning a sequence of reactions that can produce the molecule in a lab. Microsoft Research’s RetroChimera addresses that task by combining two retrosynthesis models with complementary strengths. In blinded evaluations reported in a Nature paper, PhD chemists preferred its proposed disconnections to competing model outputs and held-out literature reactions. A separate route-planning test accepted its complete routes for nine of ten challenging targets.
Why route planning explodes
Retrosynthesis starts with a target molecule and works backward, repeatedly breaking it into simpler precursors until every required input is available as a purchasable building block. Each step can produce many plausible disconnections, causing the search tree to expand quickly across a vast range of chemical transformations.
Common failure modes include missing rare but useful reactions, producing implausible precursors, and losing accuracy on molecules that differ from the training data. A planner must also rank chemically possible reactions according to practical value, since a weak step can invalidate an otherwise credible multistep route.
Two models cover each other’s gaps
R-SMILES 2 is a Transformer-based generative model that predicts precursor molecules directly as SMILES strings, a text notation for molecular structures. Its open-ended generation supports reactions involving substantial structural changes and patterns learned from data, though it can also propose unsupported precursor sets.
NeuralLoc represents the target molecule and reaction templates as graphs. A reaction template records a reusable bond-change pattern extracted from known reactions. The model selects a template and predicts where to apply it, grounding its output in observed chemistry while limiting coverage to transformations represented in the template library.
R-SMILES 2 performs particularly well when a reaction changes much of the molecular structure. NeuralLoc handles localized changes and low-precedence reactions, meaning transformations that appear infrequently in the training data. RetroChimera combines their ranked predictions instead of relying on either model alone.
Ranked votes settle disagreements
Each submodel returns a ranked list of predicted reactant sets. RetroChimera uses a learning-to-rank system that assigns a learned vote according to the source model and a prediction’s position in that list. When both models propose the same reactant set, their votes are added before the candidates receive a final ranking.
The same voting framework can accept additional proposal sources. The paper identifies reaction-database lookups and human-in-the-loop queries as possible inputs, allowing a deployment to combine model predictions with institutional data or expert feedback.
Blind reviews and a nine-of-ten result
In blinded single-step evaluations, reviewers did not know which system produced each proposed disconnection. For complex molecules, the chemists preferred RetroChimera’s suggestions to outputs from its constituent models, established baselines, and literature-derived reactions in the held-out test set.
Route-level evaluation applied a stricter criterion because every step had to pass expert review. Across ten challenging targets, the systems produced the following results:
| System | Accepted routes |
|---|---|
| RetroChimera | 9 of 10 |
| De novo model | 5 of 10 |
| Editing model | 4 of 10 |
| NeuralSym baseline | 2 of 10 |
The route result covers ten selected targets and does not estimate performance across all chemical space. The study reports expert judgments rather than laboratory execution, so every proposed route still requires feasibility review and experimental validation.
Transfer helps, data still constrains
The paper reports three capabilities relevant to industrial use:
- Rare-reaction recall: recovering useful reaction types that occur infrequently in training data.
- Zero-shot transfer: applying the model to a new dataset without additional training.
- Private-data adaptation: fine-tuning on proprietary reaction records that an organization cannot publish.
Generated predictions can still be chemically implausible, and template-based components inherit the coverage limits of their libraries. Published reaction datasets also favor successful, well-documented experiments, leaving failed reactions and niche transformations underrepresented.
From prediction to workflow
Microsoft released RetroChimera under the MIT license in its GitHub repository. The model is also available through Microsoft Foundry, and the repository provides checkpoint-access instructions.
RetroChimera supplies precursor predictions that existing multistep search frameworks can combine into complete routes. Teams building design-make-test workflows must still provide the surrounding search process, chemist review, and experimental testing before a proposed synthesis reaches the lab.
Comments
