We all have encountered similar issues when using Glimmer, GenMark, Galaxy and so on ! My try- Maker [http://derringer.genetics.utah.edu/cgi-bin/MWAS/maker.cgi] as well. Honestly, unless you specify the "minimum ORF length at some 100s aa" etc as cut-off, you have little chances to ameliorate such problems ! That is the prime reasons why all the gene prediction softwares produce "numbers that are NOT comparable" ! Also, this is far from practicable to have a very short TM-protein so as to be comparable to another predicted with lengths of almost 4 times longer ? ! If those TM families could be parsed to "length basis depending on their known annotations from previously reported orthologs- it would ease the job", won't it ? Thanks.

Yes, depending on the "relatedness" of the taxa ! Depends on how you "infer" those relatedness- say all the 3 species are strictly belonging to a single Family or Sub-family ! But, what if they are "allopolyploid" in origin ? What if the "relatedness" on taxa as previously inferred are merely "taxonomic/ phenetic system based" and not true-relatives ? Lower group's gene structures and higher groups gene structures in a kingdom would significantly differ. Well, agreed with your approach (and thats what MOST of us do too !) if you compare Chimps, Human and Gorilla as training sets for "Bonobos or "Orangutans" ! Case specific and worth trying until a "numerical consensus is made" !