Configuration: Borzoi (128-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 89.1% Accuracypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (128-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (128-BP)', column 'Accuracy' (PDF page 6 of 11) Configuration: Borzoi (128-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 92.9% Precisionpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (128-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (128-BP)', column 'PPV' (PDF page 6 of 11) Configuration: Borzoi (128-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 39 (79.6%) Recallpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (128-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (128-BP)', column 'Sensitivity' (PDF page 6 of 11) Configuration: Borzoi (128-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 76 (96.2%) Specificitypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (128-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (128-BP)', column 'Specificity' (PDF page 6 of 11) Configuration: Borzoi (128-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 3 (3.8%) False positive rate (type I error: false positives as a share of the benign set)percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (128-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (128-BP)', column 'Type I error' (PDF page 6 of 11) Configuration: Borzoi (128-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 10 (20.4%) Proportion (type II error: false negatives as a share of the splicing set (false-negative rate))percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (128-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (128-BP)', column 'Type II error' (PDF page 6 of 11) Configuration: Borzoi (32-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 91.4% Accuracypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (32-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (32-BP)', column 'Accuracy' (PDF page 6 of 11) Configuration: Borzoi (32-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 93.2% Precisionpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (32-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (32-BP)', column 'PPV' (PDF page 6 of 11) Configuration: Borzoi (32-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 41 (83.7%) Recallpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (32-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (32-BP)', column 'Sensitivity' (PDF page 6 of 11) Configuration: Borzoi (32-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 76 (96.2%) Specificitypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (32-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (32-BP)', column 'Specificity' (PDF page 6 of 11) Configuration: Borzoi (32-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 3 (3.8%) False positive rate (type I error: false positives as a share of the benign set)percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (32-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (32-BP)', column 'Type I error' (PDF page 6 of 11) Configuration: Borzoi (32-bp) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 8 (16.3%) Proportion (type II error: false negatives as a share of the splicing set (false-negative rate))percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Borzoi (32-bp) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'BORZOI (32-BP)', column 'Type II error' (PDF page 6 of 11) Configuration: Evo2 40B (human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 92.2% Accuracypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (HU. IDENT.)', column 'Accuracy' (PDF page 6 of 11) Configuration: Evo2 40B (human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 86.8% Precisionpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (HU. IDENT.)', column 'PPV' (PDF page 6 of 11) Configuration: Evo2 40B (human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 46 (93.9%) Recallpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (HU. IDENT.)', column 'Sensitivity' (PDF page 6 of 11) Configuration: Evo2 40B (human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 72 (91.1%) Specificitypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (HU. IDENT.)', column 'Specificity' (PDF page 6 of 11) Configuration: Evo2 40B (human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 7 (8.9%) False positive rate (type I error: false positives as a share of the benign set)percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (HU. IDENT.)', column 'Type I error' (PDF page 6 of 11) Configuration: Evo2 40B (human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 3 (6.1%) Proportion (type II error: false negatives as a share of the splicing set (false-negative rate))percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (HU. IDENT.)', column 'Type II error' (PDF page 6 of 11) Configuration: Evo2 40B (without human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 93% Accuracypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (without human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (W/O HU. IDENT.)', column 'Accuracy' (PDF page 6 of 11) Configuration: Evo2 40B (without human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 88.5% Precisionpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (without human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (W/O HU. IDENT.)', column 'PPV' (PDF page 6 of 11) Configuration: Evo2 40B (without human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 46 (93.9%) Recallpercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (without human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (W/O HU. IDENT.)', column 'Sensitivity' (PDF page 6 of 11) Configuration: Evo2 40B (without human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 73 (92.4%) Specificitypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (without human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (W/O HU. IDENT.)', column 'Specificity' (PDF page 6 of 11) Configuration: Evo2 40B (without human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 6 (7.6%) False positive rate (type I error: false positives as a share of the benign set)percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (without human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (W/O HU. IDENT.)', column 'Type I error' (PDF page 6 of 11) Configuration: Evo2 40B (without human identifier) (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 3 (6.1%) Proportion (type II error: false negatives as a share of the splicing set (false-negative rate))percent · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Evo2 40B (without human identifier) on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'EVO2 40B (W/O HU. IDENT.)', column 'Type II error' (PDF page 6 of 11) Configuration: Genos 10B (Alavi 2026) Protocol: Splice-defect classification of OPA1 variants with thresholds tuned on the same sets (Alavi Table III, evaluation sets) Dataset: OPA1 LOVD evaluation sets (49 splicing, 79 benign) 77.3% Accuracypercent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The AlphaGenome rows of Tables III, IV and the extended OPA1 positive calls come from scripts that pass a 0-based position to AlphaGenome's 1-based genome.Variant, so each variant was applied one base from its true position. The author labelled the report a bug on 2026-10-20 and said rerunning with the 1-based position substantially improved accuracy. The AlphaGenome results are disputed until the rerun is published; the other models' rows are not affected.Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III and Table IV, rows 'ALPHAGENOME (SPLICE SCORER)', 'ALPHAGENOME (JUNCTION SCORER)' and 'ALPHAGENOME (EITHER MODULE)'; scripts 2c8_alphagenome_splicescorer.py lines 98 and 120 and 2c8_alphagenome_junctionscorer.py lines 89 and 110 at commit 3c0c6cc; https://github.com/marcel-2c8/Probing-genomic-foundation-models-2026/issues/2
Methods, coverage and source Genos 10B on OPA1 LOVD evaluation sets (49 splicing, 79 benign)
rna-splicing-20261011-alavi2026-protocol-opa1-evaluation
Aggregation: Not reported
Probing Genomic Foundation Models with Splice-Variant Perturbations (Alavi, bioRxiv 2026) · Table III, row 'GENOS 10B', column 'Accuracy' (PDF page 6 of 11)