Back in the mid-1990s, when I worked in Chemical Discovery at DuPont Ag, we ran a brainstorming session on what would be needed to use algorithms (what later became known as AI) to devise chemical structures that would hit active sites, have desirable soil and phloem mobility properties, appropriate volatility, degradation characteristics, favorable tox across a broad range of off-target species, etc., etc. And, be manufacturable at a reasonable cost.
This sort of approach was already pretty advanced in the pharmaceuticals field, but we aggies were dead set on reinventing the wheel, our way.
This led eventually to a meeting with our in-house patent attorneys to see if patents derived from such computer-aided discovery projects could be patentable as such. One specific example we tossed them was a potential patent application based not on shared structural characteristics, as most such composition of matter patents are, but on the basis of about a half-dozen somewhat vaguely related chemical structures (chemical cousins instead of chemical siblings) that were active herbicides and that fit a specified, computer-derived structure/activity relationship correlation.
The "unobvious characteristics" upon which the IP were to be based would be the finding that the diverse structures all fit this correlation, which was computer derived, and all had the same activity profile in the field. In Ag, field tests are really field tests!
If granted, that patent would provide coverage for any new cousin or even distantly related structures to be discovered that fit the SAR correlation. A yummy can of worms.
The attorneys' immediate response once our presentation was finished: Who would be the inventor?
There had to be at least one human inventor, someone well versed in the art, who had made the inventive step.
They emphasized to us that, legally, a computer could not make an inventive step.
And our precious idea died at that meeting, in the mid-90s.
From the linked article, it looks like you can't keep a good idea down forever!