Reverse Deep Learning Engineering in Pharma

If workable, quite a coup.  Its back to the idea that it is all patterns.

A.I. finds non-infringing ways to copy drugs pharma spends billions developing  By Luke Dormehl  in DigitalTrends

Drug companies spend billions developing and protecting their trademark pharmaceuticals. Could artificial intelligence be about to shake things up? In a breakthrough development, researchers have demonstrated an A.I. which can find new methods for producing existing drugs in a way that doesn’t infringe on existing patents.

Called Chematica, the software platform does something called “retrosynthesis,” similar to the kind of reverse engineering that takes place when an engineer dissects an existing product to see how it works. In the case of Chematica, this process is based on a deep knowledge of how chemical interactions take place. It has around 70,000 synthetic chemistry “rules” coded into its system, along with thousands of additional auxiliary rules prescribing when particular reactions occur and with which molecules they’re compatible. An algorithm then inspects the massive number of possible reaction sequences in order to find another way to the same finish line.

“They effectively walk on enormous trees of synthetic possibilities, so it is a graph search problem they are trying to solve,” Bartosz Grzybowski at the Ulsan National Institute of Science and Technology (UNIST) in South Korea, told Digital Trends. “Akin to chess, every synthetic move these algorithms make is evaluated by scoring functions, which we developed over the years to tell the program whether it is navigating in the right synthetic direction.’” .... "