Computational planning of the synthesis of complex natural products

Abstract

Teaching computers to plan multistep organic syntheses has been a challenge for over 50 years1–7. Since early pioneering contributions, including programs such as LHASA1,7 (with reaction choices at each step made by human operator), the field has progressed greatly and there are now multiple software platforms6,8–13 capable of completely autonomous planning. Still, these programs ‘think’ only one step at a time and have so far been limited to relatively simple targets whose syntheses could, arguably, be designed by human chemists within minutes and without computer’s help. To date, no algorithm has been able to design plausible routes to complex natural products for which significantly more far-sighted, multi-step planning is necessary14,15 and for which one cannot rely on closely related literature precedents. Here we demonstrate that such route choices are possible, provided that the machine’s knowledge of organic chemistry and data-based artificial intelligence routines are augmented with causal relationships16,17, allowing it to strategize over multiple synthetic steps. With these improvements, results of a Turing-like test administered to synthesis experts indicate that the routes designed by computer become largely indistinguishable from those designed by humans. Three computer-designed syntheses of natural products were successfully validated in the lab. Taken together, these results indicate that automated synthetic planning at an expert level is finally becoming feasible, pending continued improvements to the reaction-knowledge base and further code optimization