They already do. QSAR has been a thing for a long time. The limiting factor is almost always the data. The publicly available SAR data in chembl / pubchem is trash, and why would you spend time and money designing and generating a good training dataset for a model when you could just be optimizing the compound?
Keep in mind that the dataset generated during optimization and the one you would prefer to train a model are quite different.
Also, in practice optimizing a lead compound is about more than just potency. You are also worried about things like toxicity, bioavailability, off target effects (selectivity), synthesis routes / yields (manufacturability), etc. And functional readouts (the assays you run) are not consistent across diseases / indications / targets / compounds.
That being said I am almost certain someone has used Bayesian optimization for an assist.
IMHO comparing QSAR to modern deep learning methods is sort of like comparing linear regression to modern deep learning methods: a technique that is wholly unsuitable for the sort of problems that deep learning is good at.
While I generally agree the data quality is poor, let's be clear: by exploting the data in PDB and Uniprot, Deepmind "solved protein structure prediction". There is a lot more value in the data that's out there, just mostly underutilized.
> How can we further improve leveraging AI beyond discovery?
Poorly. AI does not extrapolate. It can explore defined interpolation spaces with an outcome artifact in input areas where there is little data, but typically cannot extrapolate outside of convex hulls, provided classes, or other hard boundaries very well.
In effect, you have to be crafty and clever about the space an AI is to search within.
Do you think we will come up with AIs that can model potency? How can we further improve leveraging AI beyond discovery?