RetroChimera is an AI-driven retrosynthesis model that merges two complementary inductive biases through a learning-based ensembling approach to enhance synthesis planning.

Zero-shot transfer and fine-tuning on internal pharma datasets demonstrate strong generalization in real-world settings.

The system aligns with expert chemists' expectations, delivering fully accepted reaction sequences for most benchmark molecules and beating other models in several cases.

The article previews Nature content with access limitations and institutional options for readers.

Authors and affiliations span Microsoft Research AI for Science, Novartis, Cambridge, Jagiellonian University, GSK, and other institutions, reflecting broad collaboration.

Implications include faster, potentially cheaper AI-driven synthesis planning for real-world research, with applicability to small-molecule therapeutics and materials science.

Validated in Nature and intended to move from benchmark testing to real-world discovery, integrating into drug development workflows to speed design-to-lab synthesis.

The study uses pairwise and pointwise evaluations, finding chemists prefer RetroChimera’s predictions over published routes and other AI models.

The project fits within Microsoft’s AI-for-science program, aiming to accelerate discovery across chemistry, materials science, and drug development.

Ongoing work seeks to improve automated retrosynthesis and assess the model in practical discovery settings beyond controlled benchmarks.

The model addresses failure modes of prior AI approaches by targeting rare but important reactions and avoiding hallucinated or misaligned predictions.

Experimental results show RetroChimera outperforms leading baselines across data scales and remains robust to distribution shifts and unseen data.