I believe the first part of this is already in use. Drug companies can tell a computer "we need a chemical that will bind to this receptor. The computer will look through the possibilities and spot out some likely candidates. The drug companies then only have to test 10 or so likely drugs as opposed to looking at random for a suitable drug.

I'd expect that future iterations would incorporate modeling to guess at side effects. It won't be perfect (especially not at first), but it would give drug researchers a better head start on which compounds would give the biggest bang with the least side effects.