Genomic heterogeneity inflates the performance of variant pathogenicity predictions

Recent studies have reported unprecedented accuracy predicting pathogenic variants across the genome, including in noncoding regions, using large AI models trained on vast genomic data. We present a comprehensive evaluation of these frontier models, showing that performance is inflated by differences in the prevalence of pathogenic variants across genomic contexts. We identify the best-performing models for each variant type and establish a benchmark to guide future progress.