A graph neural network trained on structure-to-odor data predicted the smell of never-before-smelled molecules more accurately on average than an individual human panelist, effectively passing an 'odor Turing test.'

10:26new at the time kind of neural network called a graph neural network um which was specialized for chemistry. Um it's actually very related to the transformer. It's kind of been subsumed by the transformer in in the intervening years. And we were able to predict what things smell like very well. In fact, so well that we said why don't we set up an odor touring test. Let's go find

10:48molecules that nobody's ever smelled before. Some have never been made before, like nature has not seen them. Let's predict what they smell like ahead of time and keep our predictions secret. Let's go get those molecules physically, send them to another location. I have a great collaborator that I worked with on this. His name is Joel Mainland at Monell. Let's train people to smell and describe smell. just takes it doesn't

11:08take a ton of training, maybe like eight hours to do okay at like, hey, this smells grassy or phenolic or vanilla or cucumber, you know, and maybe 50 terms uh you can be trained on. Let's double blind um have these people smell and describe completely new molecules and then compare how well this panel of people does to first of all some individual panelist because one person is always worse than the average of the

11:34panel. So let's compare that to our model. So the question is basically you know the odor touring test is if you want to make your panel better would you rather add another person or would you rather add the predictions of a model right um and it turned out our model predictions were better than any one individual panelist on average in the panel meaning we've passed a not during

11:57test like our model predictions were human quality which was pretty cool. Um that was the first thing and what we did with the neural network is we cracked it open and we looked at what's called the embedding layer which is a a part of the neural network that uh basically turns the inputs into a vector that's that is the map right and then that map is what

12:17we kind of can slice up in regions and use for classification. So this region of the map is vanilla this region of the map is redberry etc. Um without that you actually can't do that classification problem. So that embedding turned out if you do the engineering right it just kind of needs to be around 300 dimensions to work really well which is like suspicious but you know just