I'm not sure I understand why the poor sentence must be mauled so relentlessly in your final approach, but it's fine with me if it works for you.

I note that the approach you use does not seem to take account of longest versus shortest matches:  'tor SET6' can never match because  'tor' precedes it in the ordered alternation. Perhaps this is your intent, but be aware that as the code stands, longest-shortest matching behavior depends on the order in which phrases appear in the phrase list. (This is touched on in paragraph 5 of Re^3: phrase match.) See example below.

I also note there is still no provision for a 'sentence' ending in a period, although again, perhaps this contingency will never arise. Example also below.

>perl -wMstrict -le
"my @phrases = (
   'kinase i', 'hib', 'tor', 'tor SET6', 'SET6',
   'p16(INK4A)', 'cell', 'MAP', 'H1 inhibitor',
   'foo bar', 'foo', 'bar',
   );
 for (my $i=0;$i<=$#phrases;$i++) {
     $phrases[$i]=~s/\s+/  /g;
 }
 my $phrases_re = join '|', map { quotemeta } @phrases;
 for my $sentence (@ARGV) {
   print '------------------';
   print $sentence;
   $sentence=~s/\s+/  /g; $sentence=' '.$sentence.' ';
   $sentence =~ s/\s($phrases_re)\s/ \x23$1\x23 /g;
   $sentence=~s/\s+/ /g;
   $sentence =~ s/^\s+|\s+$//g;
   print $sentence;
   }
"
"kinase inhibitor SET6 MAP H1 inhibitor activates p16(INK4A) in cell-w
+all"
"tor tor SET6 SET6"  "SET6 tor SET6"  "tor tor SET6 SET6."
"foo bar bar"  "foo foo bar bar"
------------------
kinase inhibitor SET6 MAP H1 inhibitor activates p16(INK4A) in cell-wa
+ll
kinase inhibitor #SET6# #MAP# #H1 inhibitor# activates #p16(INK4A)# in
+ cell-wall

------------------
tor tor SET6 SET6
#tor# #tor# #SET6# #SET6#
------------------
SET6 tor SET6
#SET6# #tor# #SET6#
------------------
tor tor SET6 SET6.
#tor# #tor# #SET6# SET6.
------------------
foo bar bar
#foo bar# #bar#
------------------
foo foo bar bar
#foo# #foo bar# #bar#
[download]