Superpower Daily

Superpower Daily

Your reading space

Sign in
New York Sets November Registration for Frontier AI DevelopersAI Panel Publishes Warning on Losing Control of Agents After Op...xAI Releases Grok 4.7 With Longer Reasoning at Grok 4.6 PricesMicrosoft Research Open-Sources RetroChimera for Chemical Synth...Google Starts Googlebook Preorders, Putting Android Phone Apps...WIRED Review Finds Memory Limits Apple’s M6 Mac Mini for AI Age...New York Sets November Registration for Frontier AI DevelopersAI Panel Publishes Warning on Losing Control of Agents After Op...xAI Releases Grok 4.7 With Longer Reasoning at Grok 4.6 PricesMicrosoft Research Open-Sources RetroChimera for Chemical Synth...Google Starts Googlebook Preorders, Putting Android Phone Apps...WIRED Review Finds Memory Limits Apple’s M6 Mac Mini for AI Age...
Swipe

Microsoft Research Open-Sources RetroChimera for Chemical Synthesis Planning

The newly published system ranks free-form molecular predictions against template-guided ones, a design intended to handle both familiar and rare reactions.

By 2 min read
Microsoft Research Open-Sources RetroChimera for Chemical Synthesis Planning
Microsoft Research Open-Sources RetroChimera for Chemical Synthesis Planning

Listen to this story

The audio brief

About 1:37
0:001:37
Read transcript
Microsoft Research has open-sourced RetroChimera, a system that plans how to make a molecule by working backward from the final target. In an evaluation of ten difficult synthesis targets, Microsoft says RetroChimera found successful routes for nine. Its de novo model succeeded on five, while the baseline NeuralSym succeeded on two. The system is designed to combine two different ways of reasoning about chemistry. R-SMILES 2, a Transformer model, predicts precursor molecules directly. That gives it freedom to recognize patterns beyond a fixed catalog, but it can also produce hallucinated structures. NeuralLoc takes the more constrained route: a graph neural network chooses reaction templates and decides where to apply them. That makes its suggestions more grounded in training data, but limits it to reactions represented in its template library. RetroChimera does not simply pick one model. Both rank possible reactant sets, and an ensemble combines those rankings with rank-dependent voting. It gives extra weight when both models suggest the same reaction. Microsoft says that helps the system perform across both common and rare reaction classes. In blind assessments, PhD-level chemists also preferred its individual predictions over those from earlier models for reactions reported in the literature. The important caveat is scale: the full-route comparison covered only ten challenging targets. The code and weights are now available under the MIT license on GitHub, with access through Microsoft Foundry, so the next test is whether the advantage holds across the much wider range of molecules and reactions researchers actually need.

Story brief

3 key points

Microsoft Research released RetroChimera, an MIT-licensed retrosynthesis system that combines a free-form precursor generator with a template-based reaction planner. In Microsoft’s evaluation, it produced successful routes for nine of 10 difficult targets, versus five for the de novo model and two for NeuralSym. Chemists also preferred its individual predictions in blind tests. The open weights and Microsoft Foundry...

  1. 01

    RetroChimera combines R-SMILES 2’s free-form generation with NeuralLoc’s template-grounded reaction selection.

  2. 02

    Microsoft’s 10-target evaluation found success on nine targets, compared with five for the de novo model and two for NeuralSym.

  3. 03

    The ensemble uses rank-dependent voting and gives extra weight when both component models suggest the same reaction.

Microsoft Research has open-sourced RetroChimera, an AI system for working backward from a target molecule to propose routes for making it. The Nature-published work joins two models with different limits, then learns how to rank their suggestions into one synthesis-planning output.

Retrosynthesis begins with a desired molecule and repeatedly breaks it into simpler chemical precursors, aiming to reach building blocks that can be bought. The difficulty is not just proposing one plausible reaction: a useful route requires choosing a sequence of disconnections from a very large set of possibilities.

RetroChimera combines R-SMILES 2, a Transformer model that predicts precursor molecules directly, with NeuralLoc, a graph neural network that selects reaction templates and where to apply them. Microsoft says the first approach can learn patterns freely but can generate hallucinated outputs; the second is grounded in templates from training data but is limited by what its template library covers.

Rather than selecting one method, the system has both models rank possible reactant sets. It combines those rankings with learned, rank-dependent votes, adding weight when both models propose the same reaction. Microsoft says this lets the ensemble approximately match the stronger sub-model across reaction classes, including common and rare ones.

Diagram showing a target molecule processed by two prediction models before learned voting ranks their outputs.
RetroChimera combines ranked predictions from its two component models. Source: microsoft.com.

Microsoft also reports that PhD-level chemists preferred RetroChimera’s individual reaction predictions over preceding models and recorded literature reactions in blind tests. Those are promising results, but the full-route comparison covers ten challenging targets, so researchers can now test how the released system performs on the molecules and reaction types relevant to their own work.

RetroChimera’s implementation and weights are available on GitHub under the MIT license, and the system is also accessible through Microsoft Foundry. The release gives researchers access to the same ensemble design behind the reported evaluations.

Sources

  1. microsoft.comImproving synthesis prediction of small molecules at scale with RetroChimera - Microsoft Research

Loading discussion...

YOUR READING SPACE

Notifications