nAPOGEE: A machine-learning platform for clinically actionable pathogenicity assessment of all mitochondrial noncoding variants

Abstract

Mitochondrial noncoding variants, particularly those in tRNA and rRNA genes, pose significant challenges for clinical interpretation due to heteroplasmy, broad phenotypic heterogeneity where symptoms can overlap with other conditions, and the limited availability of well-established genotype-phenotype correlations. Despite their central role in mitochondrial translation, these variants have remained largely unexplored by the existing variant-effect predictors. Here, we present nAPOGEE, a novel machine-learning framework specifically designed to assess the pathogenicity of all possible single-nucleotide variants in human mitochondrial noncoding RNAs. nAPOGEE integrates two specialized predictors: tAPOGEE, which outperforms existing tools for tRNAs, and rAPOGEE, the first dedicated classifier for mitochondrial rRNA variants. Using curated training datasets, phylogenetic conservation metrics, secondary structure modeling, RNA-specific embeddings, and thermodynamic features, nAPOGEE provides biologically interpretable predictions and posterior probabilities aligned with the ACMG/AMP guidelines. Applied to both curated variant sets and population-scale data, nAPOGEE revealed consistent spatial correlation of the predicted pathogenicity, reflecting underlying structural and evolutionary constraints. This study addresses a longstanding gap in mitochondrial genomics and offers a clinically applicable tool for variant prioritization, reclassification, and research into mitochondrial disease mechanisms.