draft-sb

Near a datacenter in Mecklenburg County, Virginia, more than 230 acres are protected from development. There are eight new acres of wetlands as a habitat for frogs, fish and bugs, more than three miles of walking trails and restored streams that flow into lakes.

"I think this is a huge benefit," says Alex Gottschalk, county administrator. "Our lakes and waterways are pretty much our signature selling point. People want to enjoy them, whether that's locals or tourists coming in."

Researchers from Microsoft Research and collaborators including GSK and Novartis, two global companies at the forefront of developing new medicines, have developed an AI model called RetroChimera to help address that challenge. A new study in Nature details how the system performs across public and proprietary chemistry data and how closely its suggestions align with expert chemists' expectations.

The work builds on Microsoft's longstanding research into using AI to accelerate scientific discovery, including advances in chemistry, materials science and drug development.

How it works

Retrosynthesis is a technique that starts with a desired target molecule and works backward, step by step, to break it down into simpler, commercially available building blocks. The task is often compared to strategic board games such as chess and Go, but its decision space is vastly larger and more complex.

For decades, this immense combinatorial complexity led researchers to believe that retrosynthesis could not be automated reliably. But recent advances in machine learning are now challenging this assumption.

As described in the Nature study, RetroChimera, paired with a search algorithm, can now propose promising synthesis pathways for target molecules, demonstrating that AI can perform retrosynthetic planning at a level that increasingly complements expert human decision-making.

Rather than relying on a single AI model, RetroChimera combines predictions from multiple models with complementary strengths. This allows the system to draw on different modeling approaches when evaluating possible synthesis routes.

The researchers developed the approach after analyzing common shortcomings in existing retrosynthesis systems, including difficulties incorporating less frequent but strategically important reactions and a tendency to generate inaccurate predictions, which have so far previously made automated synthesis planning of more complex molecules unviable.

Why it matters

One challenge in pharmaceutical drug discovery and other industries is that companies often work with proprietary chemistry data that differs from the public datasets used to train many AI systems.

The researchers showed that the pre-trained RetroChimera model can be readily adapted to GSK's internal data and its proprietary chemistry data, suggesting the approach could help researchers apply AI-driven synthesis planning to practical drug discovery challenges rather than only benchmark datasets and academic research tasks.

The finding could make the technology faster and less expensive to customize for use in other real-world research environments, including small-molecule therapeutics, materials science and fine chemical development.

When expert chemists evaluated proposed pathways for 10 molecules selected to benchmark RetroChimera against other models, the new model produced a fully accepted sequence of reactions for nine, compared with two to five for other models. And when given a choice between RetroChimera's top suggestion and previo