Science2 publishers3 min readPublished
Chemists in blind tests preferred RetroChimera's disconnections to literature reactions
Microsoft published its retrosynthesis model RetroChimera in Nature and released the weights. The blind test behind the headline claim graded individual disconnections, a narrower question than whether a whole route survives a lab.
The Scientist · Science desk
What happened
- Microsoft says its retrosynthesis model RetroChimera, which automatically proposes synthesis routes for a target molecule, has been published in the journal Nature.
- The paper covers the architecture plus validation studies on recall of rare reaction types and on zero-shot transfer and fine-tuning on proprietary datasets.
- Microsoft has open-sourced both the implementation and the weights, saying it hopes the release helps researchers develop medicinally relevant molecules and advanced materials.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability Because the weights are downloadable, a company whose reaction records cannot leave its own network can fine-tune on them locally instead of sending structures to someone else's endpoint.
- decision Any group publishing single-step retrosynthesis results now has a peer-reviewed comparator it can run itself, and a reviewer can reasonably ask why it was left out of the table.
- constraint The endpoint measured is agreement with expert judgement, so a group adopting the model still owes itself bench validation before treating a preferred disconnection as a route that works.
- contradiction Microsoft's own account both claims the ensemble beats either sub-model alone and describes it as approximately matching the better sub-model per reaction class, which is a claim about consistency and not about a higher ceiling.
RetroChimera's two halves fail in different directions. R-SMILES 2 is a Transformer that writes precursor molecules straight out of the target structure, which lets it pick up reaction patterns from data alone, and Microsoft says its unconstrained generation also makes it prone to hallucination [6]. NeuralLoc encodes the molecule and the reaction templates as graphs, selects a template and predicts where on the molecule it applies; its outputs are grounded in patterns extracted from training data, and it is more constrained when a reaction is not covered by the template library [7].
The division of labour shows up by reaction type. R-SMILES 2 does particularly well where a lot of the molecule changes over the course of the reaction, NeuralLoc on reactions of low precedence and on more localized changes [8]. The combination is a vote: each model assigns a learned, rank-dependent vote to every proposed reactant set, and the votes add when both models propose the same reaction [9].
Microsoft's summary says the ensemble produces better predictions than either model alone [3]. Its description of the result is narrower, and more interesting: by learning how much to trust each model at each rank, RetroChimera approximately matches the better-performing sub-model across reaction classes [10]. In my view that property is worth having, because a planner does not know in advance which reaction class a target will demand. Microsoft lists rare-reaction recall, robustness beyond the training distribution and alignment with chemists' expectations as the standing failure modes in existing systems [12].
The preference result is about single steps. In blind tests, PhD-level chemists preferred RetroChimera's individual reaction predictions over those of preceding models and over recorded literature reactions [4], and Microsoft frames the gain as predictions that better align with chemists' judgment [13]. The company's own account splits the problem in two: choosing among immediate disconnections, and the high-level strategic thinking needed to reach an end-to-end synthesis plan [16]. A step-level preference says a chemist likes the move. Whether the finished route runs at the bench is a separate question. The blog post does not report how many chemists took part or how large the preference was [15].
The implementation and weights are public [2], and the paper reports successful zero-shot transfer and fine-tuning on proprietary datasets alongside tests of rare reaction recall [5]. Those two facts together are what a medicinal chemistry group actually needs from a release. Reaction records that cannot leave the building can be used for fine-tuning. The comparison against internal tooling can be run on internal targets. Microsoft's stated reason for the work is cost. It says planning how to make molecules is still largely manual, time-consuming and costly, and that synthesis is a significant driver of drug development costs [11].
What to watch
- Whether an independent group reproduces the blind preference with its own chemists, its own targets and a published effect size.
- Whether route-level evaluations, not single disconnections, show the same preference, and whether any proposed route is run at the bench.
- Whether the reported zero-shot transfer to proprietary datasets holds when other companies fine-tune the released weights on their own reaction records.