According to Microsoft, researchers at Microsoft Research and collaborators including GSK and Novartis have developed an AI model called RetroChimera to help plan how to make molecules in a lab. A study published in Nature describes how the system performs across public and proprietary chemistry data and how closely its suggestions align with expert chemists' expectations. RetroChimera targets retrosynthesis, starting from a desired target molecule and working backward to simpler, commercially available building blocks, a task Microsoft compares to chess or Go but with a vastly larger decision space. Rather than relying on a single model, RetroChimera combines predictions from multiple models with complementary strengths, paired with a search algorithm to propose promising synthesis pathways. The researchers say the pre-trained model can be readily adapted to GSK's internal and proprietary chemistry data, suggesting AI-driven synthesis planning can reach practical drug discovery rather than only benchmark datasets.
Our takeOur take: Retrosynthesis is a good test of whether AI helps science or just writes about it, because the output, an actual route to a molecule, can be checked in a lab. Combining several models instead of trusting one is a sensible hedge against the field's known failure mode of confident but wrong routes. The most credible part is the adaptation to a pharma company's private data, since that is where most real chemistry lives; the caveat is that aligning with expert chemists' expectations measures plausibility, not whether the routes actually work.