Detecting species-site dependencies in large multiple sequence alignments