pull a z key, dilate it slightly (by a pixel or so) and multiply it by the alpha (or edgedetect and diallate and multiply by coverage), use this as key for the comp.
if you have multiple overlapping alphas you need to convert them to coverage first then you need to do a edge detect and use the diallated edge for coverage calculation. (you should render coverage mr does this btw)
(or my favorite render out the oversampling into a file and pull the key form there and the reconstruct in your 2d comp, this works well with maya sw and renderman but pulling it out of MR is slightly trickyer but it works just remeber to bring the sample ranges to negative and bring the image output larger, it should take equal amount of time to render since its the same operation just ses a bit more memory during render thats all)