Thread: One of the biggest problems I have with models like GANs that try to find molecular structures that fit a few endpoints like ADME and affinity is that they are trying to solve inverse problems.
An inverse problem is one where you are back-calculating a vast set of solutions based on a few endpoint values. This is a massively underdetermined situation since the number of solutions that fit only a few endpoints is huge, and uniquely narrowing it down is very hard.
In 2010 the brilliant Sydney Brenner wrote a fantastic critique of systems biology in which he said that inverse problems can only be solved by injecting a priori data and constraining solution space: Watson and Crick's experimentally guided DNA structure was a classic example. 
That is why I firmly believe that most approaches for using tools like GANs for finding novel structures will fail unless guided by a priori knowledge; the dilemma is that the latter might render them accurate but less useful because they will then discover incremental things.
In this sense, any tool like generative model faces the same problem that many computational chemistry tools face; when there's little data, they are useless because they are inaccurate. When there's a lot of data, they can be useless because they are superfluous.
The trick is to find that middle ground where there's enough data to make a dent in the inverse problem wall, but not enough for other simpler approaches and human intuitions to find good solutions. Sadly such situations are very rare, and identifying them beforehand very hard.
I am not enough of a mathematician to figure out if this would work, but I wonder if anyone has used the theory of "sloppy models" to approach molecular design and to find rigid parameter combinations that can usefully fit the data lassp.cornell.edu/sethna/Sloppy/…
