What is the hardest thing you think scientists need to do in a lab? Organic chemistry may not be the first thing that comes to mind, but multi-step organic syntheses are easily ranked a top challenge, even among experienced chemists. Nevertheless, computer scientists surprised us again with artificial intelligence (AI) which, despite having less chemistry experience than the average high-schooler, could prescribe recipes with success.

Researchers from Germany and China trained a neural network on 12.4 million single-step reactions. The network could learn from these data to predict multi-step reactions that produce complex molecules, without being taught any chemistry rules beforehand. In a double-blind test of this algorithm, eight out of its nine suggestions worked properly in the lab.

We are hopeful this new friend will bring fruits to the field and keep chemists happy for a long time. Meanwhile, we raise exciting questions just by reflecting upon this astonishment – that an advanced calculator can do so well what seems to us, all this time, like a black magic. Certainly, an AI can have a bigger storage than human memory. But we, too, have our neural networks and have been intuiting things like how to produce target molecules. We know we have not reached our brains’ full potential, but how far off are we, and how do we improve? Are there things we can learn from how AIs learn? Can AIs inspire us with better ways to do science? We look forward to answering these questions in the not-so-distant future, maybe by ourselves, maybe with the help of AIs.

Managing Correspondent:

Hechen Ren

Original Research Article:

Planning chemical syntheses with deep neural networks and symbolic AI – Nature

Media Coverage:

Need to make a molecule? Ask this AI for instructions – Nature     

AI can invent new ways to create complex molecules – MIT Technology Review