Improving synthesis prediction of small molecules at scale with RetroChimera

Synthesis planning remains a bottleneck in drug discovery and materials design: computational methods can propose many candidate molecules, but figuring out how to make them is still slow, expensive, and expertise-heavy. Microsoft Research reports RetroChimera in Nature, combining two models with a learned re-ranking step so the ensemble can outperform either alone.

In blind tests described in the post, PhD-level chemists preferred RetroChimera’s individual reaction suggestions over preceding models and even over recorded literature reactions. On ten challenging multistep targets, RetroChimera produced accepted complete routes for nine, versus five for the de novo model, four for an editing model, and two for NeuralSym. The team open-sources implementation and weights under MIT and offers access through Microsoft Foundry.

We’re for tools that shorten the design-make-test loop when the evidence is concrete. Retrosynthesis is a combinatorial planning problem—closer to chess or Go in branching, but with a far larger, less obvious move set—and most prior systems struggle on rare reactions, out-of-distribution molecules, and alignment with what chemists actually trust.

RetroChimera’s bet is complementarity, not a single bigger model. R-SMILES 2 is a Transformer that invents precursors freely and handles large structural changes but can hallucinate. NeuralLoc is a graph network grounded in reaction templates: more reliable locally, weaker when the template library doesn’t cover the chemistry. Learning how to vote across ranks is the interesting engineering move—it lets the system lean on whichever specialist is stronger for a reaction class.

The chemist preference results and the nine-of-ten route acceptances are the strongest claims in the blog. They’re still expert judgment on proposed routes, not a randomized trial of wet-lab success rates, cycle time, or cost. Preference over literature reactions is striking; it doesn’t by itself prove the routes will work on the bench.

I’m watching how open weights change who can stress-test this. Microsoft says zero-shot transfer and fine-tuning on proprietary datasets worked in their studies. That’s the enterprise path: pharma keeps its reaction history private while adapting a public base model. The downside if this scales poorly supervised is confident wrong routes that look chemist-plausible—especially on rare chemistry where human review is thinnest.

Paired with lab automation, the trajectory the authors sketch is closed-loop synthesis planning and execution. Useful if ownership, validation, and failure logging stay as rigorous as the Nature evals. Harmful if “AI said synthesize this” becomes a shortcut past experimental controls.

Retrosynthesis works backward from a target molecule to purchasable building blocks. Microsoft Research positions RetroChimera as an ensemble framework rather than a single predictor, and invites the chemistry community to probe strengths and failure modes on real targets.