my $grammar = <<'__EOI__'

    {
        use strict;
        use warnings;
    }

    parse    : list /\Z/ { $item[1] }
    list     : '(' list_ele(s?) ')' { $item[2] }
    list_ele : list | TERM
    TERM     : /\w+/

__EOI__
[download]

#!/usr/bin/perl

use 5.010;

use strict;
use warnings;

use YAML;
use Regexp::Grammars;

my $str = <<'EOT';
(root
  (class living_thing
   (spec human
        (prop has_legs))
   (spec bird (prop has_wings))))
EOT

my $parser = qr{
    ^ <List> $

    <rule: List> (?: \( <[MATCH=Atom]>* \) )
    <rule: Atom> <MATCH=List> | <MATCH=Word>
    <token: Word> \w+

}x;

print Dump $/{List} if $str =~ /$parser/;

__END__
---
- root
-
  - class
  - living_thing
  -
    - spec
    - human
    -
      - prop
      - has_legs
  -
    - spec
    - bird
    -
      - prop
      - has_wings
[download]

I sense an XY Problem here. What you need to do is instead of deciding your data structure (array of arrays), you need to look through the data you're given, the questions you're going to need to answer, and then try to come up with something that works.

I have only one sample of your input and I don't know what you're going to be asked, so I can't make good suggestions. However I can say that fitting an arbitrary tree into an array of arrays is a bad idea. "OTOH is there some uniformity that you haven't discussed? For instance are the only types of nodes root, class, spec, and prop in arrangements structured like your example? Could AI::Prolog be helpful for you?

Without knowing context I can't even tell whether those are useful questions to ask!

That looks like Lisp. Perhaps Data::SExpression would be useful?