Have you played around with LAST? I have not, but it might be of interest. Other than that I think you’re right, DIAMOND might be the best bet for large scale annotations. I try to avoid the whole mess by doing highly targeted annotations whenever possible, or for metagenomes annotating assembled contigs and then mapping.

I haven’t done any really systematic testing, but I have re-run a 78Mb plant transcript set which we previously annotated to Uniprot, using blast. We previously found hits at 1E-10 for 61% of the sequences, whereas with Diamond vs. the NR database I get only 41%, and that is at 1E-3 (same result at 1E-5). Of course it took 78 minutes, compared to a week with Blast.
I also tried the –sensitive setting, which only raised the percentage to 42%, while taking 3x as long.

It makes sense to me that the longer seeds of Diamond would be rather less sensitive than blast, and also less than programs like Ublast, or my own Fastannot program, which use 5mer seeds. However for the moment Diamond seems like basically the only game in town for aligning to NR. Am I missing anything?

I agree that there hasn’t been a lot of discussion, although they have netted 41 citations – not bad for a paper that’s been out less than a year. I also haven’t done a real side by side comparison yet with other programs of greater sensitivity. How much have you played around with the different options/parameters?