SpliceLoc-ML: A Location-Stratified Machine Learning Model for Non-Canonical Splice Variant Pathogenicity Prediction and Classification, with Application to Long QT Syndrome

Abstract

Non-canonical splice variants that occur outside the ±1/±2 GT-AG canonical splice site positions, represent a clinically important but computationally neglected class of genomic variation. Although splicing-disrupting variants account for approximately one-third of all disease-causing variants, the ACMG/AMP classification framework offers only weak computational evidence criteria for this class. Consequently, a large portion of pathogenic variants in these regions are excluded completely from analysis pipelines or in some cases classified as VUS, leaving patients undiagnosed despite carrying a disease-causing mutation that could be identified. To address this gap, this study develops SpliceLoc-ML, a location-stratified ensemble meta-classifier that predicts the pathogenicity of non-canonical splice variants and maps its output directly onto the ACMG/AMP five-tier classification. Three separate Random Forest models were trained, for exonic, near-splice (3–20 bp), and deep intronic (>20 bp) variants on 27,994 variants: 6,828 experimentally validated pathogenic variants from SpliceVarDB and 21,166 benign controls from gnomAD v4.1, integrating scores from SpliceAI, Pangolin, MaxEntScan, and CADD-PHRED with the GC content and polypyrimidine tract density. Thresholds were anchored to the Tavtigian Bayesian posterior probability framework with the lower VUS boundary derived from the training data using Youden's J statistic. All three models exceeded pre-defined performance thresholds, with AUROC values of 0.969, 0.975, and 0.936 and MCC values of 0.837, 0.853, and 0.811 for exonic, near-splice, and deep intronic variants respectively, outperforming every individual tool evaluated. SpliceLoc-ML was then applied to 4,349 non-canonical splice variants from Long QT Syndrome genes. The number of VUS decreased by 56.6% (from 489 in ClinVar to 212), and the model placed 54 variants in the Likely Pathogenic or Pathogenic tiers. These are computational assignments and need functional and clinical confirmation. These results establish SpliceLoc-ML as a reliable framework that connects computational splice variant prediction directly to clinical variant reporting.