Nature-published retrosynthesis AI 'RetroChimera' released… Demonstrates superior performance over existing models in expert evaluation

Microsoft has released the implementation code and weights of RetroChimera, a retrosynthesis AI model published in the Nature journal, under the MIT license. This model is expected to contribute to improving the efficiency of new drug development and smart material design by automatically proposing complex molecular synthesis routes.

Complementary Model Ensemble Architecture

RetroChimera is a framework that combines two models with different strengths in an ensemble manner.

• R-SMILES 2: A Transformer-based de-novo model, strong in reactions involving large changes during the reaction process, but with a risk of hallucination.
• NeuralLoc: A Graph Neural Network (GNN)-based model, demonstrating excellent performance in low precedence reactions and reactions involving localized changes.

By integrating the prediction results of the two models using a learned re-ranking strategy, it proposes more accurate and reliable synthesis routes than a single model.

Expert Evaluation and Performance

In a blind test involving PhD-level chemists, individual reaction predictions by RetroChimera were preferred over existing models and reactions recorded in the literature. In particular, it found successful multi-step synthesis routes for 9 out of 10 difficult targets, demonstrating superior performance compared to existing models such as NeuralSym (2 successes).