TL;DR: Microsoft combines two different AI chemists into one planner that beats either alone

TL;DR: Microsoft Research and collaborators with GSK and Novartis reported RetroChimera in Nature on September 21, 2026, an ensemble that merges a free-form Transformer with a template-grounded graph network to propose laboratory synthesis routes.

Microsoft Research said its new retrosynthesis system proposes complete, high-quality routes from a target molecule back to purchasable starting materials, and PhD-level chemists preferred its single-step suggestions over earlier models and literature records in blind comparisons, according to the Microsoft Research blog titled "Improving synthesis prediction of small molecules at scale with RetroChimera."

The result matters because synthesis planning remains manual, slow and expensive, even as software designs ever more candidate drugs and materials. RetroChimera does not rely on one modeling style. It pairs R-SMILES 2, which writes precursor structures directly, with NeuralLoc, which picks known reaction templates and locates where to use them, then learns how to rank their combined proposals, according to the same Microsoft Research blog.

Across the validation described in the two Microsoft reports, the ensemble recalled uncommon but strategically useful reaction types, transferred to company data, and completed nine of ten difficult multi-step targets, compared with two to five for other methods. The team released implementation and weights to let outside laboratories test and extend the approach, with researchers planning evaluation in live discovery programs rather than benchmarks alone, according to the Microsoft Source article titled "RetroChimera: New research advances AI-assisted molecule synthesis."

Why making molecules is the bottleneck

Custom molecules now underpin work in modern medicine, smart materials and sustainable agriculture, yet turning a drawing on a screen into powder in a vial remains laborious. Chemical synthesis builds a new compound from simpler blocks through sequential laboratory operations, and that construction phase drives a large share of development cost and delay, according to the Microsoft Research blog.

The mismatch has widened as computation improved. Modeling tools can enumerate huge libraries of novel candidates, but each candidate still needs a practical preparation method. Without a workable route, a promising design stays theoretical, according to the same Microsoft Research blog.

Microsoft frames RetroChimera as a direct response to that gap. Instead of accelerating design alone, the project automates route suggestion, aiming to shorten the interval between identifying an interesting structure and testing it physically, according to the Microsoft Source article.

The industrial context sharpens the need. Pharmaceutical groups typically guard internal reaction collections that look different from public training sets. A planner trained only on open data may stumble on the chemistry a company actually runs, according to the Microsoft Source article.

Collaborators in the study included GSK and Novartis, described as two global firms at the forefront of new-medicine development. Their involvement allowed testing on both open benchmarks and restricted corporate collections, a combination intended to show relevance beyond academic leaderboards, according to the Microsoft Source article.

The broader Microsoft effort sits within a longer program to apply artificial intelligence to scientific discovery, spanning chemistry, materials science and drug development, according to the same Microsoft Source article.

What retrosynthesis means in practice

Retrosynthesis inverts the normal laboratory direction. A chemist names the final target, then repeatedly splits it into progressively simpler precursors until the chain ends at compounds that can be ordered from a supplier, according to the Microsoft Research blog.

Each split is a disconnection — a proposed single reaction run in reverse. A full plan links several such disconnections into a tree. Branches represent alternatives considered; one highlighted path becomes the recommended preparation, with circles for molecules and squares for reactions in the team's illustration, according to the same Microsoft Research blog.

The Microsoft teams compare the task to chess and Go. A player surveys immediate moves while also judging long-term strategy toward checkmate or territory. A retrosynthesis program likewise must weigh an immediate disconnection and its consequences for the eventual route to available starting points, according to the Microsoft Research blog.

Chemistry makes the game harder. The Microsoft Research blog notes that the count of plausible transformations dwarfs the move sets of board games, and it is often unclear which transformations even apply to a particular target. The Microsoft Source article adds that for decades that combinatorial scale convinced many specialists that dependable automation was out of reach.

Current tools still struggle in three linked ways, according to the Microsoft Research blog: recalling infrequent but tactically valuable reactions, staying reliable outside the training distribution, and producing suggestions that match chemist expectations. Because of those gaps, high-quality planning has demanded deep specialist judgment, limiting scale and automation.

Paired with a search routine, RetroChimera is presented as evidence that machine learning now complements expert choice at this planning level, rather than merely ranking isolated reactions, according to the Microsoft Source article.

Two models with opposite trade-offs

RetroChimera's central design choice is to avoid betting on a single inductive bias. It houses two strong single-step predictors that fail in different ways, then learns to adjudicate between them, according to the Microsoft Research blog.

The first component, R-SMILES 2, is a Transformer-based de novo generator. Given a target structure, it emits candidate precursor molecules character by character in SMILES notation. Learning patterns straight from data gives the component latitude to capture transformations not codified in advance, according to the Microsoft Research blog.

That freedom carries a cost. Because generation is unconstrained, the Transformer can invent implausible arrangements — the hallucination problem familiar from language models, transferred to molecular strings, according to the same Microsoft Research blog.

The second component, NeuralLoc, takes a graph neural network approach. It encodes both the target molecule and candidate reaction templates as graphs, then chooses a template and predicts the site on the target where that template should operate, according to the Microsoft Research blog.

Grounding in templates extracted from training reactions makes NeuralLoc's outputs steadier and more precise. The limitation is coverage: if a useful transformation never entered the template collection, the component cannot easily reach it, according to the same Microsoft Research blog.

Performance splits along chemical lines. According to the Microsoft Research blog, R-SMILES 2 handles reactions that cause extensive restructuring particularly well, while NeuralLoc shines on low-precedence transformations and those involving confined, local edits. Instead of duplicating each other, the pair captures distinct slices of reactivity.

The researchers said they adopted the ensemble after reviewing recurring weaknesses in prior planners, especially omission of rare but decisive reactions and generation of erroneous proposals that blocked automation for elaborate targets, according to the Microsoft Source article.

How learned ranking fuses the pair

The fusion framework, shown as Figure 2 in the Microsoft Research post, starts with a target molecule fed in parallel to the sub-models. Each sub-model returns its own ordered list of possible reactant sets. A learning-to-rank layer then merges those lists into one ordering, according to the Microsoft Research blog.

Mechanically, each model contributes a learned, rank-dependent vote for every reactant set it proposes. When both models converge on the same reactant set, their votes sum, lifting consensus proposals. By tuning how much to trust each model at each rank position, the ensemble approaches the stronger contributor within each reaction category, according to the same Microsoft Research blog.

The design leaves room for more inputs. Although the reported work uses only deep-learning predictors as sources, the diagram includes dashed boxes for other possible contributors, such as lookups in reaction databases or queries to a human specialist, according to the same Microsoft Research blog.

That extensibility matters for deployment. A laboratory could in principle inject proprietary rules or curator feedback without rebuilding the core, though the published study tests only the two-model configuration, according to the same Microsoft Research blog.

The Microsoft Source article emphasizes the conceptual payoff: by drawing on multiple modeling outlooks when scoring possible routes, the system avoids locking into the blind spots of any one architecture. The researchers describe this as achieving planning that increasingly complements, rather than replaces, human judgment, according to that Microsoft Source article.

Training and inference therefore balance novelty and discipline. One arm explores beyond fixed catalogs; the other anchors suggestions in observed precedent. The ranker learns when exploration deserves priority and when grounding should win, according to the Microsoft Research blog.

What blind chemist tests showed

Validation covered architecture, recall of uncommon types, and alignment with practitioner taste. The Nature paper, as summarized by Microsoft, includes extensive checks plus zero-shot transfer and fine-tuning on proprietary collections, according to the Microsoft Research blog.

The most visual test involved ten demanding targets requiring multi-step routes. Reviewers scored individual steps and then judged each complete route as accepted or rejected. RetroChimera delivered an accepted full sequence for nine targets, according to the Microsoft Research blog and the Microsoft Source article.

Comparison systems trailed on the same ten targets. The de novo model alone succeeded on five, an editing-based model on four, and a baseline called NeuralSym on two, according to the Microsoft Research blog. The Microsoft Source article summarizes the rival range as two to five successes, underscoring the gap between single-style planners and the ensemble.

Step-level preference told a similar story. In blind assessments, doctoral-level chemists favored RetroChimera's individual reaction predictions over those from preceding models and over recorded literature reactions, according to the Microsoft Research blog.

A second head-to-head sharpened that finding. Nine PhD-level organic chemists drawn from Microsoft and major pharmaceutical firms were offered RetroChimera's leading suggestion alongside a previously documented preparation for the same molecule. They selected the RetroChimera proposal about 64% of the time, according to the Microsoft Source article.

The authors present that 64% figure not as proof of universal superiority but as evidence of chemist alignment — suggestions that specialists find credible enough to prefer over established precedent in a majority of direct choices, according to the same Microsoft Source article.

Scale remains a caveat noted by outside coverage and implicit in the study design: ten full-route cases demonstrate strategic behavior but cannot represent the full diversity of medicinal and materials chemistry. Open release is positioned as the remedy, inviting broader stress-testing, according to the Microsoft Research blog.

Moving from public benchmarks to company libraries

A central claim concerns portability to private data. Many corporate archives use reaction distributions, solvents, and building-block preferences absent from open sets, causing public-trained planners to degrade upon transfer, according to the Microsoft Source article.

The team reported that a pre-trained RetroChimera model adapted readily to GSK internal data and to its proprietary chemistry data. That adaptability suggests synthesis planning could address live drug-discovery tasks, not only curated benchmarks and academic exercises, according to the same Microsoft Source article.

Fine-tuning and zero-shot transfer were both examined, according to the Microsoft Research blog. Zero-shot means applying the model without additional training on the new collection; fine-tuning means brief further training on that collection. Success in both modes would lower customization cost for new settings.

If that portability holds generally, adaptation could become faster and cheaper across small-molecule therapeutics, materials science and fine-chemical development, according to the Microsoft Source article. Those fields share the same underlying need: convert a designed structure into an orderable, executable protocol.

The finding also speaks to robustness beyond the training distribution, one of the three weaknesses flagged for earlier systems. Handling company-specific chemistry without collapse indicates the ensemble's voting scheme generalizes rather than memorizing benchmark quirks, according to the Microsoft Research blog.

Researchers said the next phase will move beyond benchmark evaluation into real discovery workflows to learn how the model behaves as part of the drug-discovery process, according to the Microsoft Source article. The stated long-term aim is modest in tone: make RetroChimera one reliable instrument in the kit that carries a design from concept to bench-ready method, according to the same Microsoft Source article.

Open code and what comes next

Microsoft said it is open-sourcing RetroChimera's implementation and weights to help investigators accelerate work on medicinally relevant compounds and advanced materials, according to the Microsoft Research blog. Availability through standard code and model channels is intended to let academic and industrial groups reproduce results, probe failure modes, and build extensions.

Authorship reflects the cross-sector scope. The Microsoft Research blog credits Felix Pultar, John Gardner, Guoqing Liu and Marwin Segler as authors of the announcement, while the Nature study itself lists contributors from Microsoft Research, Novartis, the University of Cambridge, Jagiellonian University and other partners.

For United States laboratories, the practical takeaway is procedural. A group can download the ensemble, test it on internal targets alongside existing planners, and measure acceptance by its own chemists before committing synthesis resources. The ten-target and 64% preference numbers provide reference points, not guarantees, for that internal evaluation.

The scientific takeaway is architectural. Rather than chasing a single ever-larger reaction model, RetroChimera shows that pairing divergent representations — sequence generation and graph-grounded template selection — and explicitly learning when to trust each rank can lift both recall and acceptability, according to the Microsoft Research blog.

What remains is field validation at scale: diverse targets, costed routes, and laboratory execution. The team plans to study performance inside active discovery settings, where building-block availability, yield, safety and throughput decide whether a proposed path is truly viable, according to the Microsoft Source article.

Sources referenced