The output errts dataset should not be used for registration, because it no longer looks like a brain. The AFNI processing should be able to register it to the template space of your choice, then registration would not be needed in MELODIC. FEAT and FLIRT are likely not needed here.
It might be good to find out (from the FSL folks) what sorts of preprocessing are acceptable before running MELODIC. I expect they would not want censoring done, for example (I am not sure whether you are doing that).
- rick