++wfsp. For uniformity, I would store *every* value of %gp as an arrayref, rather than need conditional logic on every use of a value.

Here is my quick refactor, with ++hbm's test data (padded to 5 columns):

use strict;
use warnings;
my %gp = (
    G11  => [qw( Galpha-11  )],
    G12  => [qw( Galpha-12  )],
    G13  => [qw( Galpha-13  )],
    G14  => [qw( Galpha-14  )],
    G15  => [qw( Galpha-15  )],
    G16  => [qw( Galpha-16  )],
    Gs   => [qw( Galpha-s   )],
    Gz   => [qw( Galpha-z   )],
    Golf => [qw( Galpha-olf )],
    Go   => [qw( Galpha-o   )],
    Gq   => [qw( Galpha-q   )],

    Gi   => [qw( Galpha-i1 Galpha-i2 Galpha-i3 )],
    Gt   => [qw( Galpha-t1 Galpha-t2 Galpha-t3 )],
);

my $re = qr{
  \A
     ( [^\t]* \t [^\t]* )
  \t ( [^\t]*           )
  \t ( [^\t]* \t     .* )
  \z
}msx;

LINE:
while ( my $line = <DATA> ) {
    chomp $line;

    my ( $col_1_2, $col_3, $col_rest ) = ( $line =~ /$re/ ) or do {
        warn "Line $. did not match: '$line'";
        next LINE;
    };

    $col_3 =~ tr{ }{}d;

    my $aref = $gp{$col_3} or do {
        # No replacement needed? Is line still wanted?
        warn "Did not find '$col_3' in %gp";
        print $line, "\n";
        next LINE;
    };

    for my $val ( @{ $aref } ) {
        print join( "\t", $col_1_2, $val, $col_rest ), "\n";
    }
}
__DATA__
biologist    xargon    Gi    question    col5
perl    monks    G11    answer?    col5
[download]