There is the chance the computational model is wrong.
You only hear about the human trials that go wrong.
We still use both to weed out treatments, for both ethical and cost reasons. Better to errantly eliminate a working drug occasionally than to let a lot of unhelpful stuff through.