AI driven retrosynthesis and cooking #chmoinformatics? #cooking #pytorch

‘AI drug discovery’ is hot topics in pharma in these days. There are lots of publications about not only de novo drug design but also synthetic planning. Synthetic planning is called retro synthesis which is method to decompose target molecule to small fragments which can purchase from supplier.

For AI driven retro synthesis, input data is target molecule structure. Then the AI deradate compound to small fragments with learned chemical transformations.

BTW, do you like cooking? I like it but not good at cooking. I often search recipe with google. I think cooling and organic synthesis is same because they have recipe(experimental procedure) and it is required know how. So I wonder that are there any research about de novo cooking.

I found it !!!!

Researchers in Facebook developed inverse cooling AL with the state of the art machine learning technique.

You can find the details and code from following URL.
https://arxiv.org/abs/1812.06164
https://github.com/facebookresearch/inversecooking

The code detects photo of food and then predicts ingredient and procedure with RNN.

I felt the process is same as retro synthesis with AI.

So I had interested the code and tried to use it.

I borrowed the code from original repository and changed food photo.

The result was below.

I tried curry rice, sushi and bread. The model failed to predict curry rice and sushi.

AI missed salmon to carrot. lol

State of the art Resnet is used the algorithm but it is not perfect. It’s same as AI driven retro synthesis I think.

Current DL based model is not perfect so sometime it makes disappointing us but it has room of improvement. Near the future, we might eat food which is designed by AI. ;)