A new study from Microsoft Research in collaboration with the Broad Institute compared 400 AI models and the results were surprising