atonality is way too much of a “macro” effect to really be noticed by an mp3 encoder. the psychoacoustic model is not even close to discerning what key something is in. (intuitively: if it did, then mp3 would be awful for anything but music, whereas it is really just mediocre.)
even though the most apparent use case for mp3 is popular music, it’s used for other things (notably, movie audio tracks) so it’s important to have a general encoder which doesn’t make that many assumptions beyond the purely psycho-acoustic (and ideally, anatomical/physical).
it’s more likely that the instruments used in those pieces are harder to compress (or even that they are obscure noisy recordings, as space_monkey notes).
