Splice predictors and genomic foundation models on OPA1 and multi-gene splice variants (Alavi 2026)
Proxy evidenceIndependent study: Reported in a study whose authors did not develop the evaluated tool.Accuracy (higher is better), with the other metrics below each value. The best value in each column is highlighted.
| Tool | OPA1 evaluation sets (thresholds tuned here) | Independent 65-gene test set |
|---|---|---|
| 96.1% (best in column) | 68.3% | |
SPLICEAImachine learning | 95.3% | 94.5% |
| 93.0% | 85.2% | |
| 91.4% | 86.9% | |
| 89.1% | 71.6% | |
| 77.3% | 61.7% |
R recall, P precision, Sp specificity.
Caveats and method
What each column measures:
- OPA1 evaluation sets (thresholds tuned here): Accuracy, PPV, specificity, sensitivity and error rates of SpliceAI, AlphaGenome (version 2 rerun), Borzoi and the genomic foundation models NTv3 650M, Evo2 40B and Genos 10B on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
- Independent 65-gene test set: Accuracy, PPV, specificity, sensitivity and error rates of SpliceAI, AlphaGenome (version 2 rerun), Borzoi and the genomic foundation models NTv3 650M, Evo2 40B and Genos 10B on Multi-gene spliceogenic test set (76 spliceogenic, 107 benign, 65 genes)
- In-sample: thresholds were optimised on these 128 variants, so the values are optimistic.
- Positives are LOVD entries annotated as splicing defects; the assay behind each annotation is not stated.
- Thresholds were set by one-dimensional grid search on the OPA1 evaluation sets, optimising accuracy and specificity (Methods B).
- Single-author preprint; no uncertainty intervals are printed for the tabulated metrics.
- Foundation-model scores are zero-shot log-likelihood ratios with author-chosen context windows, aggregation and strand rules (Methods F to H).
- Preprint (bioRxiv versions 1 and 2), not peer reviewed; version 2 changed only the AlphaGenome rows.
- AlphaGenome's three version 1 configurations stay excluded over the 1-base coordinate offset in their scripts (upstream issue #2). Its four configurations from version 2 of the preprint (posted 2026-10-31) were rerun with the corrected position and are included.
- Spliceogenic positives span canonical ±1,2 (26 of 76), near-splice intronic positions from +3 to +7, -3 to -7 and -13 (35) and exonic substitutions (15), by the paper's Supplemental Table VI, although its text calls them canonical splice-site variants; benign controls are ClinVar intronic substitutions at ±3 to ±10 from 46 of the 64 gene symbols the table lists.
- Positives were confirmed by RNA studies in the cited sources (references 13 and 14); this preprint does not restate the assay per variant.
- Genos 10B ClinVar is fine-tuned on ClinVar labels and the benign controls are ClinVar variants; training overlap is not tested in the source.
Protocols: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets), Spliceogenic versus benign classification on an independent 65-gene set with OPA1-derived thresholds (Alavi Table IV)