As oshalla pointed out, pack 'b*' pads the string to the nearest byte.

In order to know how much you need to trim off after you've unpack'd it, you need to also remember how many binary bits need trimming off the unpacked binary; 0, 2, 4, 6.

You can do that by prefixing the packed data with an extra byte that contains that information:

Update: My math was completely wrong and only worked for the example through blind luck. I've cleaned that up and wrapped the conversions in subroutines:

my %charmap = (
    A => '00',
    C => '01',
    G => '10',
    T => '11',
);
sub packACGT {
    ( my $in = shift ) =~ s/(.)/$charmap{$1}/ge;
    print length $in, ':', length( $in ) % 8;
    pack 'Cb*', ( 0, 6, 4, 2 )[ length( $in ) % 8 / 2 ], $in;
}

my %digmap = (
    '00'    => "A",
    '01'    => "C",
    '10'    => "G",
    '11'    => "T",
);

sub unpackACGT {
    my( $len, $out ) = unpack 'Cb*', shift;
    print $len;
    substr( $out, -$len ) = '' if $len;
    $out =~ s/(..)/$digmap{$1}/ge;
    return $out;
}

my $string = $ARGV[0] ||  'GATTACCCC';

print length( $string ), ':', $string;
my $compressed = packACGT( $string );

print length( $compressed ), ':', "'$compressed'";

my $decompressed = unpackACGT( $compressed );

print length( $decompressed ), ':', $decompressed;
[download]

It would be a nice enhancement to pack and unpack to allow the n/A* syntax to deal with this.

So that pack 'c/b*', ... and unpack 'c/b*', ... would work.