RetroChimera retrosynthesis model outperforms baselines and wins chemist preference

September 21, 2026 at 09:24 PM1 min

This digest was compiled by AI from multiple sources — links to the originals are below.

Microsoft Research and partners introduced RetroChimera, a retrosynthesis model that outperforms leading baselines across data scales. Organic chemists preferred its predictions over published reference reactions and other AI models. The model demonstrated zero-shot transfer and fine-tuning on internal datasets from Novartis and GSK.

Key Facts

RetroChimera was developed by Microsoft Research AI for Science with contributors from Novartis, GSK, University of Cambridge, Jagiellonian University, and Bergische Universität Wuppertal.
The model outperforms leading baselines across several orders of magnitude in data scale and learns from very small numbers of examples per reaction class.
Organic chemists preferred RetroChimera predictions over published reference reactions and other AI models in both pairwise and pointwise setups.
RetroChimera demonstrated zero-shot transfer and fine-tuning on internal datasets from Novartis and GSK, showing robust generalization under distribution shift.

Model Architecture

RetroChimera integrates two newly developed components with complementary inductive biases through a novel learning-based ensembling strategy. The model was built by Microsoft Research AI for Science in Cambridge, UK, with equal contributions from Krzysztof Maziarz, Guoqing Liu, and Felix Pultar. The research team analyzed failure modes of current AI models, including struggles with less frequent but strategically critical reactions and hallucinated predictions misaligned with chemists' expectations.

Experimental Validation

Experiments across several orders of magnitude in data scale showed RetroChimera outperforming leading baselines and demonstrating robustness outside the training data. In pairwise and pointwise evaluations, organic chemists preferred RetroChimera predictions over published reference reactions and other AI models. The model achieved zero-shot transfer and fine-tuning on internal datasets from Novartis Biomedical Research in Basel and GSK in Stevenage.

Pharmaceutical Industry Application

RetroChimera showed robust generalization under distribution shift when applied to internal datasets from two major pharmaceutical companies. The work demonstrates the viability of deep learning for accurate synthesis prediction in increasingly challenging regimes.