RetroChimera AI model plans molecule synthesis routes

Designing promising drug molecules has become increasingly sophisticated, but translating those designs into practical laboratory synthesis remains one of chemistry's most complex challenges. A new AI system called RetroChimera aims to bridge that gap by suggesting viable pathways to construct target molecules from commercially available chemical building blocks.

Developed by Microsoft Research in collaboration with pharmaceutical companies GSK and Novartis, the model represents a shift in how AI can support retrosynthetic planning—the process of working backward from a desired molecule to identify the sequence of reactions needed to create it. The research, published in Nature, demonstrates performance that aligns closely with expert chemist judgment across both public and proprietary datasets.

How the system works

RetroChimera tackles retrosynthesis by combining predictions from multiple AI models rather than relying on a single approach. This ensemble strategy allows the system to leverage different modeling strengths when evaluating potential synthesis routes.

The researchers designed this architecture after identifying common failures in existing retrosynthesis tools, including difficulty incorporating strategically important but infrequent reactions and a tendency toward inaccurate predictions that made automated planning impractical for complex molecules.

Paired with a search algorithm, RetroChimera proposes synthesis pathways by breaking down target molecules step by step into simpler components. The decision space for this task exceeds the complexity of strategic games like chess or Go, which had long led researchers to doubt whether reliable automation was achievable.

Performance against expert judgment

In benchmark testing, expert chemists evaluated proposed pathways for 10 molecules. RetroChimera produced fully accepted reaction sequences for nine of those molecules, compared with two to five for competing models.

When nine Ph.D.-level organic chemists from Microsoft and major pharmaceutical companies compared RetroChimera's top suggestions against previously documented synthesis methods for the same molecules, they preferred the AI system's approach approximately 64% of the time.

Why it matters

Pharmaceutical companies and other chemistry-focused organizations typically work with proprietary datasets that differ substantially from the public data used to train most AI systems. The researchers demonstrated that RetroChimera's pre-trained model can be adapted to GSK's internal chemistry data, suggesting the approach could scale to practical drug discovery applications rather than remaining limited to academic benchmarks.

This adaptability could reduce both the time and cost required to customize AI-driven synthesis planning for real-world research environments, including small-molecule therapeutics development, materials science, and fine chemical production. The ability to work with proprietary data addresses a significant barrier to AI adoption in industrial chemistry settings.

Next steps

Researchers plan to evaluate RetroChimera in active drug discovery workflows, moving beyond controlled benchmark testing to assess how the system performs as part of the actual development process. The goal is for the model to become a standard tool that helps researchers move efficiently from molecular design to viable laboratory synthesis pathways.

The findings were first reported by Microsoft in a Nature study. The RetroChimera code is available on GitHub.