Most unbound sites from the human genome mispredicted by mouse-trained models (x-axis), but not by human-trained models (y-axis), contain Alu repeats. For visual clarity, only 5% of windows are shown.