In case someone is interested about this domain.
There was an interesting challenge around it on AIcrowd last year.
It was called 'Learning to Smell' and hosted by popular swiss fragrance and flavour business Firmenich. It contains publicly available dataset, baselines, winner codebases, etc., to play around with.
Some fluff in this PR piece, but I love the message that molecular chemistry can now also benefit from deep learning advances, thanks to graph neural networks. Because molecules are by their nature graphs of atoms, it makes sense to learn a latent representation of a molecular entity with a GNN, and use that for various classification or prediction tasks.
Makes me wonder if the computationally expensive geometry optimization calculations using DFT etc. can be partially done by a GNN. So, pre-optimize a structure with a computationally cheaper GNN to speed up convergence.
This has been a previous area of research for Google (https://ai.googleblog.com/2017/04/predicting-properties-of-m...). It remains routine to benchmark GNNs and other molecular machine learning models on predicting quantum mechanical properties including energies (which speed up geometry optimization)
is legal and political. Google has it in digital form, but they can't give it to you, sell it to you, show it to you, or do anything else with it. So the researchers had to re-digitize it. Or maybe ("The Google Books team brought the USDA dataset online.") the researchers had some special help from Google.
Thanks. I was in Google Patent Litigation, and there was a book I wanted to see. I didn't know it was valuable; if it was, there'd be no problem buying it, but most of the time, these things were false positives.
Someone in Books offered to let me come over and see it on HIS screen, because he couldn't give me the link even as a Google employee.
But I guess Sciences (or GAS, or whatever it's called), got y'all to help. Good for them (and you).
Good then: is your research reproducible? As in, do you have some code and data somewhere that can be used to verify your results?
By that, I mean to run it on the input dataset to create something that gives results similar to you GNN results for whatever testset you used (in your example, predicting the mosquito repellency of a set of molecules)
exciting direction in this largely unexplored area.
Empirically we have done a bunch of work with correlating mixtures of molecules to human perception at Volatile AI (https://volatile.ai) and the degrees of variation in mixtures are really wild - a tiny amount of something in a mixture can change the smell perception of the whole mixture.
So getting decent results will be so much harder when looking beyond single molecules
Thinking about smell a few years ago I realized that individual differences in odor perception might be due to missing smell receptors.
In color vision we have only three different kinds of cones. Color blindness due to a missing type of cone is present in a substantial portion of the population.
I believe one of my children is colorblind, and it took a long time to figure that out. With a missing smell receptor it would be even more difficult to detect the differences between people.
I'm curious how much Google's model would be improved by not aggregating the individual humans together, instead treating their identities as a part of the dataset.
If you want to have an idea of the molecular mechanisms involved, last year the structures of an insect odorant receptor in complex with two different odorants was published: https://www.nature.com/articles/s41586-021-03794-8
I’ve been wondering for years if it might be possible to use a UAV with a pretty hefty sniffer to check out some areas and detect cannabis pollen. Then it would maneuver toward the origin based on wind data and if it hits it would photo and log GPS and auto fly to base ASAP. Then Subcontract to some hardcore criminals to go rip it off and flip it. Totally risk free money!
seems like a risk to life (stealing drugs, hiring criminals, etc) when you could just move somewhere where its legal and start a legal business. And all for a fairly low value crop compared to hard drugs.
I mean it doesn't sound impossible it just sounds like it isn't worth it even if you could do it.
It was called 'Learning to Smell' and hosted by popular swiss fragrance and flavour business Firmenich. It contains publicly available dataset, baselines, winner codebases, etc., to play around with.
[1] https://www.aicrowd.com/challenges/learning-to-smell
[2] https://discourse.aicrowd.com/c/learning-to-smell/357
[3] https://www.aicrowd.com/challenges/learning-to-smell/noteboo...
Disclaimer: I have been affiliated with AIcrowd.