Abstract
Many pathogenic variants implicated in Mendelian diseases impair normal protein function, often through loss-of-function effects, while loss-of-function mutations in tumor suppressor genes commonly contribute to tumorigenesis. However, many disease-causing variants act through gain-of-function or other mechanisms that do not strictly disrupt the protein. Interpreting rare and novel variants remains a major challenge in clinical genomics, highlighting the need for computational tools informed by large, well-curated clinical datasets to reliably distinguish truly deleterious mutations from neutral variation. We developed MutAnt, a mutation meta‑annotator based on machine‑learning. It is trained on a large, clinically relevant dataset of variants using multiple variant properties, including synchronised predictions from other algorithms. MutAnt models demonstrate high F1 and ROC‑AUC scores (0.88-0.99) on hold‑out datasets and provide well‑calibrated probability scores that correlate with functional assays. MutAnt's deleteriousness predictions exhibited correlations with functional scores obtained from deep mutational scanning assays for tumor suppressor proteins BRCA1, PTEN, and p53 (ρ = 0.28-0.61), and with protein stability measurements from computational models. Moreover, MutAnt prediction scores of deleteriousness improved somatic variant calling from RNA sequencing data compared to standard approaches. MutAnt's high performance in distinguishing neutral and protein-disrupting mutations highlights its potential clinical utility in variant classification.