RetroChimera

RetroChimera is a retrosynthesis model that takes a target molecule, expressed as SMILES, and produces a ranked set of plausible chemical reactions to synthesize it. It combines several machine-learning retrosynthesis models with complementary inductive biases via a learning-based ensemble strategy, capturing both graph- and sequence-based views of molecular transformation. The system has been validated through blind expert review: PhD-level organic chemists preferred RetroChimera’s proposed routes over the ground-truth reactions on which it was trained, a strong signal of chemist-aligned predictions.

Automated retrosynthesis compresses the early phase of drug discovery and chemical manufacturing by enabling fast exploration of synthesis pathways before any wet-lab commitment. The ensemble approach improves robustness across diverse chemical space, while preserving caveats around lower-ranked predictions where hallucination risk rises and expert review remains required. Public checkpoints trained on Pistachio, USPTO-50K, and USPTO-FULL let chemists evaluate route feasibility directly during molecular design and pair naturally with property predictors and generators such as TamGen and MatterGen in Microsoft AI for Science’s chemistry portfolio.