Summary
A Nature paper introduces RetroChimera, an AI model that combines two chemically different modelling approaches to predict routes for synthesising small molecules. The researchers report stronger performance than leading baselines, favourable evaluations from organic chemists and transfer to internal pharmaceutical datasets.
A research team has introduced RetroChimera, an artificial-intelligence model designed to predict how a complex small molecule could be made from simpler chemical starting materials. In a paper published in Nature on 21 September 2026, the researchers report that the system outperformed leading baseline models, handled reactions outside its training data and produced predictions that organic chemists preferred in comparative evaluations.
The work comes from researchers at Microsoft Research AI for Science, Novartis Biomedical Research, the University of Cambridge, GSK and other institutions. The published article is an accepted early-access version with a permanent DOI and may receive further edits before the final Version of Record.
Why retrosynthesis is difficult
Retrosynthesis reverses the usual direction of a chemical reaction. Instead of starting with known reactants and predicting the product, a retrosynthesis system starts with a target molecule and proposes the reactions and precursor compounds that could be used to make it.
This is a central task in the discovery and manufacture of functional small molecules, including compounds considered during pharmaceutical research. A useful system must account for many possible transformations while prioritising routes that are chemically plausible and useful in practice.
The researchers identify two recurring problems in existing AI systems. Models can struggle with uncommon but strategically important reactions because those reactions appear infrequently in training data. They can also generate incorrect or “hallucinated” predictions that do not match the expectations of practising chemists.
These problems are linked to the way machine-learning models learn patterns from chemical data. A model may be highly capable on frequent reaction types but less reliable when the available examples are sparse or when a target molecule differs from the data it has seen.
How RetroChimera combines different models
RetroChimera is built from two newly developed components with complementary inductive biases. In machine learning, an inductive bias is the set of assumptions or preferences that guides a model towards particular patterns in data. Different biases can make models sensitive to different chemical relationships.
Rather than selecting one component as the sole predictor, the researchers combine them through a learning-based ensemble. An ensemble uses the outputs of multiple models or modelling approaches to produce a final prediction. In this case, the aim is to combine distinct strengths rather than rely on one type of chemical representation or reasoning pattern.
The study tested the system across datasets spanning several orders of magnitude in size. The researchers report that RetroChimera outperformed leading baseline systems and remained robust when evaluated outside its training distribution. They also report that it could learn useful behaviour from very small numbers of examples for particular reaction classes.
The evaluation included both pairwise and pointwise comparisons involving organic chemists. In the pairwise setting, chemists compared alternatives; in the pointwise setting, they assessed predictions individually. Across these evaluations, the chemists preferred RetroChimera’s predictions over published reference reactions and over predictions from other AI models, according to the paper.
Testing transfer to pharmaceutical datasets
The researchers also examined whether the model could transfer to data from settings different from its original training environment. They tested zero-shot transfer, in which a model is applied to a new dataset without task-specific retraining, and fine-tuning, in which it is adapted using examples from that dataset.
These tests used internal datasets from two major pharmaceutical companies. The paper reports robust generalisation under this distribution shift, suggesting that the model’s performance was not limited to the conditions represented in its original training data.
The evidence described in the article is model evaluation, expert comparison and transfer testing. The supplied article text does not give numerical benchmark scores, dataset sizes or the number of reactions in each class, and it presents the work as synthesis prediction rather than a report of a complete laboratory synthesis campaign. Its significance therefore lies in the reported modelling approach and evaluation results: combining different chemical-learning assumptions may help AI systems handle sparse, unusual and practically important reaction types while producing routes that chemists find more useful.