New AI Can Plan How to Make New Drugs — In Seconds

New medicines, smart materials and hardier crops all depend on making new molecules. But figuring out how to build a molecule — a process called synthesis — is slow, expensive and mostly done by hand by expert chemists. It's one of the biggest drivers of drug development costs. Now researchers have published a system called RetroChimera in the journal Nature that proposes step-by-step recipes for building a target molecule, and they've released the code and model weights so any lab can use it.

The idea works backward, like a chess player planning several moves ahead. RetroChimera combines two AI models with opposite strengths. One freely invents possible building blocks, which makes it creative but occasionally wrong. The other only uses reaction patterns it has seen before, so it's reliable but limited. The system learns which model to trust in each situation, and the combination beats either one alone. In blind tests, chemists with PhDs preferred its individual reaction suggestions over both earlier AI models and reactions recorded in the scientific literature.

So why should you care? Cheaper, faster molecule planning could shorten the years it takes to turn a promising compound into a medicine, and could speed up new battery materials or sustainable farming chemicals. If planning a synthesis takes weeks instead of months, that time and money saved can eventually show up as lower drug prices and treatments reaching patients sooner. The team open-sourced the model precisely so smaller labs and startups — not just giant companies — can use it.

The honest catch: RetroChimera predicts recipes, it doesn't run them. Chemists still have to test each route in a real lab, where reactions fail, scale up badly, or need unavailable ingredients. It also can't guarantee a plan is practical or affordable. Think of it as an extremely well-read research assistant that hands you a shortlist — you still have to do the experiment.