Benchmarking in silico predictors of deep intronic variant pathogenicity on a patient-derived cohort of inherited retinal dystrophy genes

Background Deep intronic variants (DIVs) are an increasingly recognized cause of inherited retinal dystrophy (IRD), accounting for a substantial fraction of cases that remain unsolved after exon-focused sequencing. Dozens of in silico tools predict variant pathogenicity or splicing impact, but they have rarely been benchmarked on patient-derived DIVs, and the most recent generation of predictors, including AlphaGenome and the phylogeny-aware genomic language model GPN-Star, has never been evaluated in this setting. Methods We assembled a cohort of 103 patient-derived candidate DIVs across 26 IRD genes (native GRCh38) and 103 benign intronic gnomAD variants (allele frequency > 0.01, all classified as benign/likely-benign in ClinVar). Ground truth was defined in two ways: 10 experimentally confirmed pathogenic DIVs and an expanded set of 36 positives obtained by adding allele-verified ClinVar pathogenic/likely-pathogenic variants. We scored all variants with seven tools spanning three methodological classes: splice-aware deep learning (SpliceAI, Pangolin, AlphaGenome, and OpenSpliceAI), general deleteriousness (CADD), and evolutionary constraint (GERP and GPN-Star). We computed AUROC, auPRC, and optimal thresholds, with a gnomAD-wide false-positive screen and a distance-matched sensitivity benchmark as robustness checks. Results Splice-aware deep-learning tools clearly outperformed all others: SpliceAI (AUROC 0.917), Pangolin (0.916), and AlphaGenome (0.902) were statistically indistinguishable, whereas OpenSpliceAI underperformed (0.842), and the constraint-based predictors GPN-Star (0.554) and GERP (0.559) performed near chance. The optimal SpliceAI threshold for DIVs was 0.06 three-fold lower than the standard recommendation of 0.2, which detected only 24 of 35 scorable positives; a formal test found no significant score-by-distance interaction; therefore, we recommend the pooled threshold, with distance-stratified estimates reported as exploratory. A screen of 4,787 common intronic gnomAD variants showed the relaxed threshold kept the population false-positive rate low (1.3% at 0.05 vs. 0.13% at 0.2), and a distance-matched sensitivity benchmark left all AUROCs and rankings unchanged. Three experimentally confirmed pathogenic DIVs scored ≤0.01 on SpliceAI, revealing a “deep intronic blind spot” for variants acting through non-splice-creation mechanisms. Consensus ranking of the 93 unlabeled patient DIVs nominated 26 novel, high-concordance candidates with no ClinVar entry. Conclusion For patient-derived IRD DIVs, splice-aware deep-learning predictors are the tools of choice, but they require deep-intronic-specific thresholds an order of magnitude below standard defaults. General constraint-based genomic language models should not be used as proxies for splice impact. We provide an actionable threshold approach and a prioritized candidate list for functional follow-up.