While this speeds up the screning process quite a lot you might as well feel like someone might just have thrown something out that might actually turn out useful and you can't backcheck without resorting to testing everthing.
When I read the headline I was expected something more along "a network that knows what processes we have and therfor could imagine what compounds we could come up with" combined with that "inverse" network that predicts weaknesses of bacteria. Then those two would be matched and we would have to figure out the missing steps how to actually build the compound coming from that "black box" prediction.
The benefit would be that we would be presented with something that could exist but we hadn't thought of because of our processes and limited resources.
While I value the speedup and cost effect I'd think we could do much better here in exploiting the "creative" and powerful numerical inverse potential of ANNs especially in this area and outsmart those evolutionary skills of bacteria that way.