One interesting issue about the protein model is how the start end points work. For proteins it is obvious that for distant homology, it needs to be local - ie can start or finish anywhere in the sequence. For protein HMMs it is less clear. If a HMM really represents a single domain then global start end points are correct. However, many times local start end points are useful.
The HMMer2 models internally carry whether this HMM is has global or local (or indeed any type) of start end policy.
However, the genewise algorithm is quite dependent on the models being global to effectively predict introns in domains, when the looping algorithm (multiple copies of the domain) is present. This is because nearly always in a local HMM, an intron can be better modelled as the end of the domain half way through and the start of a new domain half way through, further down the sequence, thus not predicting the intron. To get clean intron prediction, one needs to go to global mode. However, using global mode forces the start and end point of the model to be really correct, and in some cases (in particular some Pfam models) this makes very incorrect results on the edges of the domain. To combat this another type of start end policy is introduced - wing. This has a local start mode for the first 15 model positions and end mode for the last 15 model positions, but global in the central part of the model.
In the programs one can set four types of start end policy