I've seen a few talks on the subject. From a gene cluster, there is some decent ability to determine which domains contribute to certain functional groups in the final metabolites. It sounds like this is largely what Warp Drive Bio is trying to do. How realistic is this technology?
I would say that for type I polyketide synthases (PKS), this is possible. These proteins are variations on fatty acid synthases which have modules, each of which adds a two carbon unit as well as a chiral methyl group to an extending chain which may be reduced to give anything from a ketone, an alcohol to a completely reduced carbon chain. In the illustration below you can see these 3 options, but double bonds are also possible.
Other sorts of PKS pathways (type II and III PKSs) have proteins which are harder to predict in terms of what they produce since a single protein does the extending of the polyketide chain to cyclize it, typically into an aromatic compound. All this happens in a single protein site, so careful modelling would be required and I'd be suprised if this has been done accurately.
Note the last arrow there 'tailoring steps' - accompanying the PKS genes themselves are often a slew of acetyl and methyl transferases, additional reductases, coupling enzymes, even delivery proteins... the list goes on and on. These steps would be hard to predict because the substrate for an o-methyl transferase isn't easily determined.
I think there are several companies trying to leverage 'best guess' bioinformatics here. They are likely to get a good idea of the backbone, but maybe won't place every methyl and acetyl group correctly.
There are definitely people out there who think this could work to deliver diverse candidates to drug discovery. see Radiant Genomics as well.