2026-09-22
RetroChimera ensembles a Transformer generator and a GNN template editor. Chemists accepted 9 of 10 hard routes and preferred its top-1 over literature ~64% of the time.
Retrosynthesis works backwards from a target molecule until the pieces are buyable starting materials. One bad single-step prediction kills the whole route, so the accuracy bar is harsh. Template classifiers and SMILES sequence models have gotten good at common reactions. They still miss rare but strategic transformations, and they still invent reactions a chemist would reject on sight. Search then compounds those errors. Medicinal chemists care whether a disconnection is something they would run, not whether it matches the one reaction written in a paper.
Microsoft Research AI for Science, with Novartis, GSK, and academic collaborators, builds RetroChimera by combining two new models whose inductive biases fail in different places, then learns how to vote across their ranked lists.
NeuralLoc is the editing model. It encodes the product and each reaction template as graphs. A classification head picks templates; a localization head scores atom-to-atom matches between product and template. That second head matters when a symmetric template can be applied in more than one way. At Pistachio scale the model has about 165 million parameters and a library of 146,256 templates.
R-SMILES 2 is the de-novo model. It keeps the aligned-SMILES encoder-decoder setup from R-SMILES, then swaps in grouped-query attention, RMSNorm, and SwiGLU, and changes when beam search stops. About 66.7 million parameters on Pistachio. Editing stays closer to observed chemistry and can use a template from a handful of examples. Generation holds up better on large structural changes and far-from-train molecules, and it hallucinates more.
Ensembling happens in molecule space, not token space. Prediction rank k from model i gets a learned vote θ{i,k}; votes add when both models propose the same reactant set. The votes are fit on the validation set with a learning-to-rank loss, constrained to be decreasing and convex per model. In practice the generator gets more weight at the top ranks and the editor contributes recall further down the list.
On USPTO-50K with reaction class unknown, RetroChimera reaches 93.2% top-10 accuracy, against 91.0% for R-SMILES and 87.1% for NeuralLoc. On USPTO-FULL, top-10 is 79.5% versus 76.4% for R-SMILES. The ensemble leads for k greater than 1; top-1 still has the best model in each class within striking distance. Packing up to 10 public models into one ensemble lifts the score further, at a cost that is probably not worth paying at inference.
The main training set is Pistachio, about 3.5 times the size of USPTO-FULL. The test set is 146,393 high-quality 2024 additions whose products sit below 0.95 fingerprint similarity to anything in train. Zero-shot transfer to 10,444 internal Novartis reactions still beats both submodels under distribution shift. The Nature version also reports fine-tuning on internal data from two pharmaceutical companies; the public materials do not give the per-point fine-tune numbers.
Chemist evaluation splits "found a route" from "would run it." Nine PhD-level organic chemists ran double-blind A/B tests on about 600 pairs: RetroChimera's top-1 beat the literature reaction about 64% of the time (P<0.05). A dummy that slaps on rare templates without ranking was rejected in over 93% of those controls. The revision added about 6,000 pointwise labels to count hallucinations. On ten adversarially chosen hard targets, complete routes accepted by experts were:
| Method | Full routes accepted |
| RetroChimera | 9 / 10 |
| R-SMILES 2 | 5 / 10 |
| NeuralLoc | 4 / 10 |
| NeuralSym | 2 / 10 |
Step-level acceptance looks closer; one bad step is enough to discard a route, so the gap widens at route level.
Supplementary search on 1,784 PistachioHard targets tells a different story: R-SMILES 2 solves 93.5% by default, RetroChimera 90.7%. Tuning temperature to 0.09 or 0.5 adds about one point of solve rate and drops chemist acceptance of both steps and full routes. Solve rate by itself will happily search into hallucinations.
This is an engineering answer, not a new learning paradigm. Once the two models fail differently, learned rank votes surface the consensus, including rare reactions (one decarboxylative cross-coupling in the supplement has on the order of 10 precedents). Code is MIT-licensed; weights are on Microsoft Foundry and can sit in syntheseus as the single-step model. For a medchem group with an electronic lab notebook, the signal is that zero-shot already transfers, and fine-tuning can hug the local distribution.
The authors are clear that hallucinations remain, especially far from the training distribution, and that systematic data errors (aromatic substitutions with amine leaving groups that Pistachio recorded as valid) poison every model. Reviewers asked for wet-lab validation; the authors answered with expert labeling at scale and did not add experiments. Route-level scores still omit yield, step count, and scale-up. On the hard set R-SMILES 2 solves more targets while chemists prefer RetroChimera, so the two metrics disagree. The production ensemble uses two models; adding a strong editor such as RetroExplainer is left open. Reagents, solvents, and conditions are not predicted.