Analyzing the performance of deep learning splice prediction algorithms.

Independent benchmark of 3 SpliceAI implementations: thresholds inadequate for deep intronic variants

SpliceAI and its two open-source implementations (OpenSpliceAI and CI-SpliceAI) were independently evaluated on 6 datasets including 1,316 validated variants, 213 variants with splice assay data, 99,601 SPiP variants, 242 manually curated deep intronic pathogenic variants, and two ClinVar-derived datasets. All three deep learning algorithms outperformed a 4-tool legacy ensemble. Critical finding: for deep intronic variants, optimal thresholds are 10 times lower than standard recommendations, meaning default thresholds miss the majority of pathogenic deep intronic variants.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

This benchmark provides direct practical guidance for molecular diagnostic laboratories: it demonstrates that open-source implementations faithfully reproduce SpliceAI (>90% positional concordance), removing the barrier of restrictive licensing. Crucially, it alerts on default thresholds being inadequate for deep intronic variants — critical information for teams performing WGS reanalysis with non-coding variant searches.

Clinical impact: 2/3 · Evidence quality: 3/3 · Novelty: 1/2 · Sample size: 1/1 · Journal quality: 0/1 → Total: 7/10