Splice-altering variants account for an estimated 15–30% of genetic diseases, yet many computational predictors lose accuracy outside the canonical GT-AG dinucleotides. This limits interpretation of intronic variants of uncertain significance (VUS). MetaSplice is presented as an ensemble pathogenicity predictor designed specifically to score single-nucleotide variants across multiple non-exonic splice regions and improve clinical interpretation of intronic splice variants.
MetaSplice is positioned as a pathogenicity predictor trained on clinical labels of significance, intended to complement mechanism-specific splice-effect tools rather than replace them.
MetaSplice uses a gradient-boosted ensemble model built from 53 input features. These features integrate outputs from deep-learning splice-effect models (notably SpliceTransformer and Pangolin), measures of evolutionary conservation and gene-level constraint, and splicing-regulatory motif annotations. The ensemble structure combines mechanistic splice-effect predictions with broader sequence- and gene-level information to produce a single pathogenicity score for each candidate intronic single-nucleotide variant (SNV).
The architecture therefore leverages both modern deep-learning splice predictors and established genomic-context features to capture diverse signals associated with splice disruption and clinical pathogenicity.
MetaSplice is designed to evaluate SNVs across seven non-exonic splice regions:
By explicitly modeling these regions, MetaSplice addresses variant classes where many existing tools show reduced performance, notably within the PPT, branch-point, and deep intronic regions.
The model was trained on a large ClinVar-derived dataset of intronic SNVs. The training set comprised 381,226 intronic ClinVar SNVs, of which 35,506 were labeled pathogenic. A five-fold gene-grouped cross-validation strategy was used during development, ensuring that variants from the same gene were grouped to avoid information leakage across folds. In gene-grouped cross-validation MetaSplice achieved an area under the precision-recall curve (auPRC) of 0.995.
For independent evaluation, a temporally held-out ClinVar test set was used; model performance on that set is reported separately (see below).
On a temporally held-out ClinVar test set containing 107,933 variants, MetaSplice attained an auPRC of 0.987 (95% CI 0.985–0.989). The manuscript reports comparative performance against several established predictors on the same held-out set: SpliceAI (auPRC 0.950), CADD v1.7 (0.915), SPIDEX (0.767) and S-CAP (0.253). MetaSplice outperformed all listed comparators on the held-out data, indicating improved precision-recall characteristics for clinical pathogenicity prediction across intronic splice regions.
The paper emphasizes that MetaSplice’s gains are greatest in regions where other tools commonly underperform.
The largest relative improvements reported for MetaSplice occurred in the polypyrimidine tract (PPT), branch point, and deep intronic regions. These loci are historically challenging for splice-effect prediction because the sequence signals are more subtle or more dispersed than at canonical splice dinucleotides. By integrating motif annotations, evolutionary constraint, and multiple deep-learning splice scores, MetaSplice appears to recover signal that improves pathogenicity discrimination in these regions.
Specific numeric breakdowns by region were reported in the source; where exact per-region metrics or further stratified results are required, consult the original preprint or supplementary material for full tables and figures.
MetaSplice was applied to 47,272 ClinVar splice-region variants of uncertain significance. Using its pathogenicity scoring, MetaSplice nominated 22% of those VUS for functional follow-up. This demonstrates how a clinically trained ensemble can be used to triage intronic variants for experimental validation and potential reclassification.
The source notes this practical application as part of the tool’s intended role in variant interpretation workflows.
MetaSplice is made freely available as a Docker image, facilitating reproducible deployment and integration into analysis pipelines. The preprint is provided without peer review; users should take that into account when interpreting performance claims and before clinical application.
A competing interest is disclosed: the author, X.L., is a co-founder of Genos Bioinformatics, LLC. The manuscript and associated materials, including full methods and supplementary data, are available in the preprint and supplementary files for further inspection.
Note: This summary and rewrite strictly reflect the content reported in the cited preprint. Where the source references detailed numeric breakdowns, methodological specifics, or supplementary tables, readers should consult the original preprint for complete data and figures.