There are now lots of ways of encoding molecules. So many, in fact, that it's not really worth debating the merits of any particular method.
ECFP fingerprints shoved into a fully connected NN work surprisingly well for a large class of problems. Molecular graph convolutions (of which there are now many flavors) also work well. The field is to the point where people are doing ensembles of different encodings, and seeing what works for any particular problem.
> without resorting to empirical measurements and classification, which could yield reasonable vectors, but will take 2-5 years of a highly trained grad student's labor to obtain and put into the training corpus
Well, you're sort of touching on my last paragraph with this. The classifier, featurization, etc., usually matters less (a lot less?) than the quality of the assay data. So I agree in that respect.
What is the metric for "working well" here?
My point is a graph convolution will have a hard time distinguishing the haouamine atropisomers, because they have the same graph, but very different activities.
This sort of weird shit is the norm. Like some molecule that isn't the actual active form but needs cyp450 to be activated. But it won't be if you add a methyl group that reduces it toxicity, because that methyl blocks cyp (not your target even). But only in college student pre-phase I volunteers, who happen to be predominantly white male college students looking for beer money
In the end your corpus is going to have to fall back to experiment, which is just so much slower in chemistry than the trivially parallel "scrape every Wikipedia article, twitter tweet, and reddit comment" that you have access to for text corpora; it's more like trying to use ml to decode Chinese, if we only had 3000 Chinese documents to work with.
Model fit. Validation on a holdout set. All of the usual standards for ML performance.
> My point is a graph convolution will have a hard time distinguishing the haouamine atropisomers, because they have the same graph, but very different activities.
Isomers can be represented. The typical problem with isomers is that you don't have good data on them, because real-world laboratory solutions are rarely pure.
> Like some molecule that isn't the actual active form but needs cyp450 to be activated.
Drug metabolism is a common set of endpoints for ML models. Nothing you're describing is so mysterious or difficult that the problems "break" machine learning in this domain.
> In the end your corpus is going to have to fall back to experiment
All ML models in this space depend on experiment.
Edit: saw your other comment on GNNs. I don’t think it’s like a solved problem or anything, but I also don’t think it’s intractable. e.g. https://arxiv.org/abs/2012.00094 seems relevant
* Image space and text space are also vast, and yet we've had good success applying AI in these areas. I have yet to see a convincing argument that these spaces aren't equally large.
* It's a bit of a red herring: actual drug discovery programs are not exploring "all of chemical space". They're usually focused on "lead series" of much more constrained molecules.
* There are actually context-independent signals that can be used to generalize AI methods. The generalization is far from perfect, but it's not like every one of those 10^60 molecules are entirely different from every other molecule in the set. There are clusters and patterns and trends that can be exploited for gain -- this is what makes "medicinal chemistry" an academic field, and not merely an exercise in fortune-telling.
Personally, I think the bigger problem applying AI and ML to drug discovery is less the "vastness of chemical space" (a proposition that makes med-chemists feel secure about their jobs), and more that the datasets in drug discovery suck. There's tons of siloing of data, none of it is consistent, and you can't even depend that two assays for the same target, measured in the same lab, years apart, will yield consistent data. It's a total mess.