Discordance between a deep learning model and clinical-grade variant pathogenicity classification in a rare disease cohort.

Genetic testing is essential for diagnosing and managing clinical conditions, particularly rare Mendelian diseases. Although efforts to identify rare phenotype-associated variants have focused on protein-truncating variants, interpreting missense variants remains challenging. Deep learning algorithms excel in various biomedical tasks^1,2, yet distinguishing pathogenic from benign missense variants remains elusive^3-5. Our investigation of AlphaMissense (AM)⁵, a deep learning tool for predicting the potential functional impact of missense variants and assessing gene essentiality, reveals limitations in identifying pathogenic missense variants over 45 rare diseases, including very early onset inflammatory bowel disease. For the expert-curated pathogenic variants identified in our cohort, AM's precision was 32.9%, and recall was 57.6%. Notably, AM struggles to evaluate pathogenicity in intrinsically disordered regions (IDRs), resulting in unreliable gene-level essentiality scores for genes containing IDRs. This observation underscores ongoing challenges in clinical genetics, highlighting the need for continued refinement of computational methods in variant pathogenicity prediction.