Use RetroChimera Online

RetroChimera: Ensemble AI Retrosynthesis Model for Chemist-Aligned Single-Step Reaction Prediction

RetroChimera is a frontier retrosynthesis model that ensembles two models with complementary inductive biases, a template-based molecule-editing model (NeuralLoc) and a de novo SMILES-generation model (R-SMILES 2), and fuses their predictions with a learned ensembling framework. It was developed by researchers at Microsoft Research AI for Science and Novartis Biomedical Research, together with collaborators at the University of Cambridge and Jagiellonian University. The authors report that it is available through Azure Foundry and that a GitHub code release was being prepared.

Chemical synthesis remains a bottleneck in discovering and manufacturing small molecules. AI-based synthesis planning has progressed since 2017, but models still struggle with rarer reaction classes that matter for synthetic strategy, and they can produce hallucinated, incorrect predictions. Because multi-step search calls the single-step model recursively, errors compound and a single bad reaction invalidates a whole route.

RetroChimera addresses this by pooling diverse models rather than relying on one architecture. The paper reports that it outperforms major baselines across several orders of magnitude of data scale and splitting strategy, and that industrial organic chemists preferred its predictions over the reactions it was trained on.

How RetroChimera Combines Edit-Based and De Novo Retrosynthesis Models

The framework merges ranked output lists from several models by looking at overlaps between them. Unique reactant sets are re-ranked by a score that depends on each model's rank, and the per-model weighting is learned on a validation set with a ranking loss. The weights are constrained to be decreasing and convex, and the learned schemes can vary with rank k. Ensembling runs on CPU from saved model outputs, so it does not require re-running the underlying models.

1. NeuralLoc (editing model): a template classifier with a product encoder and a separate template encoder that reads the template as a graph. It classifies templates and localizes them on the product by computing atom-level assignment probabilities, which lets it share information between templates and resolve ambiguous template matches.

2. R-SMILES 2 (de novo model): builds on root-aligned SMILES, with an aligned input and output format and heavy augmentation. It generates reactants autoregressively and tends to generalize better far from the training data.

3. Learned voting: reactant sets proposed by both models are boosted, and rank-dependent weights decide how much each model counts. The authors note the approach could also take non-model sources such as reaction database lookups or chemist input.

RetroChimera Benchmark Results on USPTO and Pistachio

On USPTO-50K (reaction class unknown), RetroChimera reports the following top-k accuracy, against its two component models and selected baselines from the paper.

On USPTO-FULL, RetroChimera reaches 51.4% top-1, 79.5% top-10 and 85.6% top-50 accuracy. The authors state it sets a new state of the art for k greater than 1 on both USPTO datasets, raising top-10 accuracy by 1.7% and 1.6%, and that with only 10 results it matches what R-SMILES reaches with 50.

Rare Reactions, Zero-Shot Transfer and Chemist Preference

• Rare reaction classes: performance is strong for both frequent and rare templates, and the authors describe near-optimal recall on well-precedented reactions, so the model behaves like a soft reaction database.

• Zero-shot transfer: on 10,444 reactions from an internal Novartis database, with no fine-tuning and ensembling weights fixed from Pistachio, RetroChimera outperformed both of its component models.

• Multi-step search: on SimpRetro targets it reached a close to 100% solve rate at the largest time limit, and it performed best for medium-to-long search times on 800 hard Pistachio test targets.

• Expert assessment: nine PhD-level organic chemists compared predictions blind, and the paper reports that they preferred RetroChimera to the dataset ground truth in terms of quality.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with RetroChimera on Tamarind Bio

• Single-step retrosynthesis ideas: get a ranked list of candidate disconnections for a target molecule, including rarer strategic reactions that chemists may not think of immediately.

• Multi-step route planning: use the single-step predictions as the engine for retrosynthetic search toward commercially available building blocks, as the authors do with Retro*.

• Medicinal chemistry transfer: apply the model to drug-like targets, where the paper shows robust zero-shot performance on pharma internal data.

• Reaction quality triage: pair predictions with forward and feasibility models to filter implausible proposals before they reach the bench.

How to Use RetroChimera on Tamarind Bio

1. Open the tool: log in to tamarind.bio and select RetroChimera from the tool list.

2. Provide the target: enter the product molecule you want to make, typically as a SMILES string, since the model takes a product and predicts reactants.

3. Set parameters: if exposed, choose how many ranked predictions to return. The paper evaluates up to the top 50.

4. Run the job: submit it and wait for the ranked reactant sets.

5. Download results: export the predicted reactant sets and their ranks.

6. Interpret the output: top-ranked suggestions are the most reliable, and the paper analyzes agreement between the two submodels. Review each reaction with chemical judgment.

7. Use downstream: feed reactants into further retrosynthetic steps, or check them against building-block catalogs and feasibility models.

Things to Keep in Mind

• The authors note that machine learning retrosynthesis models are not free from hallucinated outputs, especially far from the training distribution.

• Systematic errors in training data can propagate, which the authors say improved data curation can mitigate.

• Reagents, solvents and reaction conditions were not predicted in the study.

• Top-1 accuracy was lower on the Novartis data than on Pistachio, which the authors attribute possibly to distribution shift.

• Pistachio is proprietary, so the headline results on it cannot be reproduced with public data alone.

Source: Maziarz, K., Liu, G., Misztela, H., Tripp, A., Li, J., Kornev, A., Gaiński, P., Hoefling, H., Fortunato, M., Gupta, R. and Segler, M. "Chemist-aligned retrosynthesis by ensembling diverse inductive bias models", arXiv:2412.05269v2 (2025), https://arxiv.org/abs/2412.05269. Model access via Azure Foundry (ai.azure.com/catalog/models/RetroChimera); the paper states a GitHub code release was in preparation.