AI tools for gene variants may score better than they should
A study in Nature Communications finds that AI models that predict disease-causing single-nucleotide variants look more accurate than they are. The problem arises when tests mix different types of variant.
The authors say benchmarks that combine variant types, such as canonical splice and 5' UTR variants, inflate the scores of these models. They propose a stratified benchmark that separates the types, and they use it to identify the most robust models. The study is peer-reviewed. Only a short summary was available, so the methods and results are not described here.