I have been testing ABySS, Velvet and SGA, but in my experience, the major problem are the chimeric sequences produced.

I would suggest using digital normalization followed by something like Velvet. However, we have not yet tried this out on metatranscriptomes. Let me know if you try it out!

The publications that I know of on metatranscriptomics first generate reference metagenomes and then map to that; see the hydrothermal vent stuff from Greg Dick's lab.

--titus