MetaSplice: pathogenicity prediction and ACMG/AMP evidence for intronic variants outside the canonical splice sites
- 1University of South Florida
Abstract
Intronic variants outside the canonical splice-site dinucleotides are difficult to interpret: splicing predictors are least accurate there, and clinical classification requires evidence about pathogenicity rather than predicted splicing alone. We developed MetaSplice, a gradient-boosted ensemble that combines SpliceTransformer and Pangolin splicing predictions with conservation, exon context and gene-level constraint, without allele frequency, to rank intronic single-nucleotide variants by pathogenicity. We trained it on 413,199 variants from the August 2025 ClinVar release and tested it on 46,112 variants classified as (likely) pathogenic or (likely) benign in September 2026 but not in August 2025. Among 40,296 non-canonical variants scored by both tools, MetaSplice reached an area under the precision-recall curve of 0.919, versus 0.757 for SpliceAI. On all 40,619 non-canonical variants, it reached 0.916, versus 0.364 for CADD. The advantage was largest at deep intronic positions, persisted when each variant was scored by a model trained without its gene, and was not clear in the polypyrimidine tract. Against experimentally assayed splicing, MetaSplice detected 92.8% of splice-altering variants at a 1.2% false-positive rate. From likelihood ratios calibrated against ClinVar, we derived a candidate rule that assigns PP3 to scores of 0.544 or higher and BP4 to scores of 0.0017 or lower, both at Supporting strength. In the test set, this rule gave PP3 to 0.20% of 37,263 benign variants, versus 4.44% under the ClinGen-recommended SpliceAI thresholds, and to 1.0% versus 31.2% of 706 formerly uncertain variants since classified benign. MetaSplice complements splicing predictors, and its candidate PP3/BP4 rule warrants independent evaluation.
Related
- In vitro pathogenicity evaluation of deep intronic variants for recessive genetic retinal diseases
- Deep-intronic variants in USH2A: the importance of model selection in variant interpretation and pathogenicity assessment
- Molecular basis of anti-CD36 alloimmunization: transcriptomic evidence linking intronic splice-site mutations to CD36 loss-of-function
- AlphaGenome informed machine learning for pathogenicity prediction of germline mutations
- Mechanism-based prediction of insertion-driven high pathogenicity avian influenza virus emergence
BibTeX
@misc{liu2026metasplice,
title = {{MetaSplice: pathogenicity prediction and ACMG/AMP evidence for intronic variants outside the canonical splice sites}},
author = {X. Liu},
howpublished = {bioRxiv},
year = {2026},
month = jul,
doi = {10.64898/2026.07.23.740393},
url = {https://www.biorxiv.org/content/10.64898/2026.07.23.740393v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - MetaSplice: pathogenicity prediction and ACMG/AMP evidence for intronic variants outside the canonical splice sites
AU - X. Liu
PY - 2026
DA - 2026/07/23
PB - bioRxiv
DO - 10.64898/2026.07.23.740393
UR - https://www.biorxiv.org/content/10.64898/2026.07.23.740393v2
AB - Intronic variants outside the canonical splice-site dinucleotides are difficult to interpret: splicing predictors are least accurate there, and clinical classification requires evidence about pathogenicity rather than predicted splicing alone. We developed MetaSplice, a gradient-boosted ensemble that combines SpliceTransformer and Pangolin splicing predictions with conservation, exon context and gene-level constraint, without allele frequency, to rank intronic single-nucleotide variants by pathogenicity. We trained it on 413,199 variants from the August 2025 ClinVar release and tested it on 46,112 variants classified as (likely) pathogenic or (likely) benign in September 2026 but not in August 2025. Among 40,296 non-canonical variants scored by both tools, MetaSplice reached an area under the precision-recall curve of 0.919, versus 0.757 for SpliceAI. On all 40,619 non-canonical variants, it reached 0.916, versus 0.364 for CADD. The advantage was largest at deep intronic positions, persisted when each variant was scored by a model trained without its gene, and was not clear in the polypyrimidine tract. Against experimentally assayed splicing, MetaSplice detected 92.8% of splice-altering variants at a 1.2% false-positive rate. From likelihood ratios calibrated against ClinVar, we derived a candidate rule that assigns PP3 to scores of 0.544 or higher and BP4 to scores of 0.0017 or lower, both at Supporting strength. In the test set, this rule gave PP3 to 0.20% of 37,263 benign variants, versus 4.44% under the ClinGen-recommended SpliceAI thresholds, and to 1.0% versus 31.2% of 706 formerly uncertain variants since classified benign. MetaSplice complements splicing predictors, and its candidate PP3/BP4 rule warrants independent evaluation.
ER -