I’ve done a bit of work on this front a few years ago. We had a number of simulators that can tell you the likelihood that something binds to a specific protein, and you’re basically just searching through all possible compounds to find things that bind to a target while minimizing their interactions with other proteins that would be known to cause problems in the human body. When we find promising candidates, we send them off to a lab to synthesize and test in vitro.

This kind of search problem isn’t exactly easy to automate mainly because naive solutions are very expensive. It’s never been a problem of outputting nonsense. The automations use the same evaluation metrics as human researchers.