There's no such thing as a "magic SAR prediction black box' whede you feed it structures and it comes back with affinity/efficacy data, unfortunately... most of the predictioms in this thread are the work of human minds drawing conclusions from already published data on drug SAR/effects.
You can find molecular dynamics or protien-ligand docking calculation software like CHARMM or GROMACS. However those programs are meant for situations where you already have a high accuracy model of the receptor in question, preferably from X-ray diffraction studies of the receptor in crystalline form. However because many receptors are typically anchored to the cell wall it can be an uphill journey to find conditions that allow a technician to both isolate the receptor in question as well as support the growth of flawless homogenous crystals of your receptor protien (plus bound ligand) that are big enough to handle and image with a cyclotron. A lot of effort is being made to derive means for recovering accurate protien structures in the solution phase just because crystallization of protiens is a black art at the best of times and seemingly impossible in the harder cases.
If you don't have the luck to have access to a good model of your target receptor, you can find a homologous receptor with a solved tertiary structure that shares a significant amount of its amino acid sequence witn your target and modify the structure accordingly, working on the assumption that the structural differences don't wildly change the shape of the receptor. I recall seeing quite a few models of different GPCRs that were based on the structure of bovine rhodopsin, for instance, because it was one of the few GPCRs people could get reasonably pure and coax into a cryatalline form.
As a last resort there is also de novo protein folding, like what Folding@Home does. However, given that protien folding is NP-complete (easy to verify a correct solution, but incredibly difficult to come up with a solution), this method takes an incredible amount of time and computational power to produce results even for protiens of relatively short length. The amount of possible configurations for each amino acid grows exponentially with every additional one so a GPCR with a few hundred to a thousand amino acids would take longer than the age of the universe to exhaust every potential conformation.
So anyway, unless you have a set of accurate models for all the receptors common to psychoactive drug effects, you're mlre likely to only be able to come up with the predicted affinity at a few receptors. Even then you'd need to mess with some pretty fancy software meant for academics and invsst some serious CPU time.