A unified framework for quantitative splicing and transcript prediction (Jaganathan et al., bioRxiv 2026)

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 1 model
• 13 configurations
• 13 protocols
• 10 datasets
• 39 evaluations
• 90 results
• 12 evidence claims

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• AlphaGenome scoring in the SpliceAI2 paper: Evaluation actor All AlphaGenome analyses were conducted at the University of Oxford by co-authors Y.K. and E.M.W., supervised by S.J.S.; the raw composite score with maximum aggregation was chosen from 40 scoring configurations as best on average across eight evaluations. (Methods 'Evaluation of competing methods', AlphaGenome paragraph; Author contributions)
• Evidence summary: patient RNA splicing validation: Summary The one recovery result already on this page is FRASER on the 119-sample Kremer fibroblast cohort, reported by its developers: a mean 85% (11 of 13) of known pathogenic splicing events recovered at a 30-sample subsample. Six published comparisons add evidence. Segarra-Casas et al. (2025), independent of the tool developers, recorded which callers reported 16 known pathogenic splicing events in clinical muscle RNA-seq from 98 samples: FRASER with |dPSI| > 0.1 reported 13, FRASER with |dPSI| > 0.3 reported 9, FRASER2 10 (and one more only with the full cohort), rMATS-turbo 7 and LeafCutterMD 4 (Table 2); the events were established partly with these callers. Segers et al. (2026), the saseR developers, ranked 12 reported disease genes in the Kremer cohort: the gene ranked within the top 10 in 9 of 12 samples for saseR-junctions, in 6 of 12 for FRASER 2.0 with the autoencoder and in 7 of 12 for FRASER 2.0 with PCA (Table 2); the 12 genes were reported in earlier analyses of the same cohort, partly with OUTRIDER and FRASER, and several are expression rather than splicing defects. Silverstein et al. (2026, preprint), a clinical research group that did not develop the splicing callers, scored them on 97 previously diagnosed neuromuscular RNA-seq samples and also counted outlier calls in true-negative samples (Table S9). On known pathogenic splicing events in muscle and fibroblast RNA-seq, rMATS-turbo detected 15 of 23 and MINTIE 8 of 23; on muscle, fibroblast and cycloheximide-treated fibroblast RNA-seq, FRASER detected 21 of 68 and LeafcutterMD none of 67; FRASER2.0, run on muscle only, detected 2 of 16. rMATS-turbo and MINTIE reported outliers in all 9 true-negative samples they were run on, FRASER and LeafcutterMD in all 21, and FRASER2.0 in 2 of 6. Denominators differ by tool, a negative sample with any outlier counts as one false positive, and the known events were chosen because they were clinically interpretable and visible in the RNA-seq, which the authors say favours clearly defined events. Drost et al. (2025), a diagnostic laboratory independent of the tool developers, scored 243 clinically ascertained variants, each tested for a splicing effect in patient cells or by exon trapping, with four DNA-based predictors. On all 243, AUROC was 0.895 for SpliceAI, 0.885 for Pangolin, 0.830 for SPiP and 0.806 for SQUIRLS (Data S1 Table S3); at literature thresholds sensitivity was 0.871, 0.858, 0.832 and 0.800 and specificity 0.750, 0.773, 0.761 and 0.591 (Table S4), and the authors' rule of requiring at least two of the four tools gave sensitivity 0.916 and specificity 0.727. Most in-house variants were sent for testing because another predictor (Alamut) flagged them, so negatives are mostly predicted-but-unconfirmed variants. Two further DNA-based comparisons add proxy evidence. Fortier et al. (2026), independent of the SpliceAI, CI-SpliceAI and OpenSpliceAI developers, report balanced accuracy on functional splice-assay variants in ABCA4 and MYBPC3 of 0.836 for SpliceAI, 0.794 for CI-SpliceAI and 0.746 for OpenSpliceAI, and on 242 deep intronic pathogenic variants with 242 common controls of 0.940, 0.890 and 0.841 (Tables 2 and 4); both sets of thresholds were tuned on the same data, so these are upper-bound estimates, and CI-SpliceAI was trained on all chromosomes. Alavi (2026, preprint) applied thresholds tuned on OPA1 variants to 76 functionally validated canonical splice-site variants and 107 ClinVar benign intronic variants from 65 genes: sensitivity and specificity were 88.2% and 99.1% for SpliceAI, 85.5% and 86.9% for pretrained NTv3 650M, 71.1% and 100% for Evo2 40B with a human identifier, 56.6% and 80.4% for AlphaGenome (either module) and 27.6% and 85.0% for Genos 10B (Table IV). Jaganathan et al. (2026, preprint), the SpliceAI2 developers at Illumina, scored four DNA-based predictors on rare variants that create a GT or AG dinucleotide in 5,435 Genomics England rare-disease participants with blood RNA-seq, against new junctions seen in the carriers' RNA (3,065 cryptic gains and 1,775,800 nulls): auPRC was 0.59 for SpliceAI2, 0.48 for SpliceAI, 0.40 for Pangolin and 0.36 for AlphaGenome, or 0.45 with AlphaGenome's raw splice site score (Fig. 4g; Methods). Only splicing visible in blood can be confirmed, so some nulls may alter splicing in other tissues, and SpliceAI2 was trained on all human genes except those on chromosomes 21 and 22. Of 23 rare variants in phenotype-matched panel genes that SpliceAI2 prioritised, 17 showed the predicted splicing change in RNA-seq; no other tool was put through this step. The paper compares no rule-based splice-site scorer. All values are transcribed from the publications, not reproduced; none is a diagnostic-yield estimate, and results from different sources measure different endpoints and must not be pooled. This summary describes evidence and does not recommend a clinical workflow. (amp-oncology-rna-20261007-source-pmc7822922: FRASER results; rna-splicing-20261009-source-segarracasas2025: Table 2, Results paragraph 3; rna-splicing-20261009-source-segers2026: Table 2; rna-splicing-20261009-source-drost2025-data-s1: Data S1 Tables S3 and S4; rna-splicing-20261009-source-drost2025: Methods; rna-splicing-20261011-silverstein2026-source-tables-s1-s10: Table S9 rows 2 to 6; rna-splicing-20261011-fortier2026-source-article: Tables 2 and 4; rna-splicing-20261011-alavi2026-source-article: Table IV; rna-splicing-20261013-spliceai2-source-article: Fig. 4g, Methods 'Evaluation of competing methods', Main 'Contribution to rare genetic disease' paragraph 4)
• Evidence summary: splicing follow-up experiments: Summary Three independent comparisons score several splice predictors against experimentally measured splicing of human SNVs; none of their authors developed the tools compared. A fourth is reported by a tool's developers. Znabu et al. (2026, preprint) scored all 27,733 labelled MFASS reporter SNVs, of which 1,050 disrupt splicing. Average precision was 0.421 for Pangolin (bootstrap 95% interval 0.386 to 0.454), 0.321 for SpliceAI, 0.317 for SpliceTransformer, 0.256 for MMSplice and 0.228 for SPANR, with AUROC 0.888, 0.819, 0.786, 0.758 and 0.748. SpliceAI was given a score of zero where it returns none. This is a different MFASS population from the matched and v2 held-out protocols and is not pooled with them. Smith and Kitzman (2023) set each tool's threshold so that it flags 10% of 500,000 random exonic and near-exonic SNVs, then measured sensitivity on saturation assays, two of which their own group produced. SpliceAI detected 0.982 of splice-disruptive variants in BRCA1 saturation genome editing, 0.885 in POU1F1 exon 2, 0.860 in WT1 exon 9, 0.539 in FAS exon 6 and 0.237 in MST1R exon 11; Pangolin detected 0.982, 0.948, 0.667, 0.487 and 0.330. Sensitivity depends strongly on the exon, precision is not measured at these thresholds, and HAL and S-Cap score only part of the variants. A curated MLH1 column is recorded but judged outside scope, because its labels include patient RNA and a pathogenicity rule. Riepe et al. (2021) tested up to 11 tools on ABCA4 and MYBPC3 variants assayed in mini- or midigenes. MCC for SpliceAI was 0.42 for ABCA4 noncanonical splice sites, 0.84 for ABCA4 deep-intronic variants and 0.35 for MYBPC3, where the Alamut 3/4 consensus reached 0.53. Those variants were chosen for testing with Alamut or MaxEntScan, and the cutoffs were tuned on the same data. Jaganathan et al. (2026, preprint), the SpliceAI2 developers at Illumina, report Spearman correlation between predicted and measured exon-inclusion change for 283,121 single-nucleotide substitutions in 599 exons of the OpenSplice reporter assay: 0.59 for SpliceAI2, 0.55 for SpliceAI, 0.53 for Pangolin and 0.49 for AlphaGenome (Extended Data Fig. 2). The comparator scores are those released with OpenSplice, while SpliceAI2 was scored by its developers, and a rank correlation over all substitutions is not precision among variants chosen for an experiment. None of these sources measures how many variants a tool would select for a new experiment and how many of those then prove splice-disruptive, so all are proxy evidence for that choice. All values are transcribed from the publications, not reproduced. Reporter and minigene endpoints do not establish patient-RNA effects or pathogenicity, and results from different sources must not be pooled. This summary describes evidence and does not recommend a method. (splicing-follow-up-20261009-source-znabu2026: Section 4 results table; splicing-follow-up-20261009-source-smith2023-table-s2: Table S2, 10% call-rate sheet; splicing-follow-up-20261009-source-smith2023: Methods; splicing-follow-up-20261009-source-riepe2021: Tables 3 to 5; rna-splicing-20261013-spliceai2-source-article: Extended Data Fig. 2 panel 'OpenSplice MPRA (ρ)', Methods 'Variant effect benchmarks')
• Relevance of Cryptic splice site activation by rare GT/AG-creating variants in Genomics England blood RNA-seq (SpliceAI2 paper, auPRC) to use case use-case-patient-rna-splicing-validation: Assessed by Cryptic splice site activation by rare GT/AG-creating variants in Genomics England blood RNA-seq (SpliceAI2 paper, auPRC) (rna-splicing-20261013-spliceai2-source-article: Fig. 4g; rna-splicing-20261013-spliceai2-source-article: Main 'Contribution to rare genetic disease' paragraph 4; rna-splicing-20261013-spliceai2-source-article: Methods 'Evaluation of competing methods' and 'Variant effect benchmarks')
• Relevance of OpenSplice MPRA delta PSI prediction for 283,121 single-nucleotide substitutions in 599 exons (SpliceAI2 paper, Spearman) to use case use-case-splicing-follow-up: Assessed by OpenSplice MPRA delta PSI prediction for 283,121 single-nucleotide substitutions in 599 exons (SpliceAI2 paper, Spearman) (rna-splicing-20261013-spliceai2-source-article: Extended Data Fig. 2 panel 'OpenSplice MPRA (ρ)'; rna-splicing-20261013-spliceai2-source-article: Fig. 1d; rna-splicing-20261013-spliceai2-source-article: Methods 'Variant effect benchmarks', OpenSplice MPRA paragraph)
• Relevance of RNA-seq confirmation of SpliceAI2-prioritised rare variants in panel genes, Genomics England rare-disease participants (SpliceAI2 paper) to use case use-case-patient-rna-splicing-validation: Assessed by RNA-seq confirmation of SpliceAI2-prioritised rare variants in panel genes, Genomics England rare-disease participants (SpliceAI2 paper) (rna-splicing-20261013-spliceai2-source-article: Main 'Contribution to rare genetic disease' paragraph 4; rna-splicing-20261013-spliceai2-source-article: Fig. 4h,i; rna-splicing-20261013-spliceai2-source-article: Extended Data Fig. 5)
• SpliceAI2 myotonic dystrophy severity contrast: Severity contrast Correlations increased between groups with greater differences in disease severity (Pearson r = 0.42; Fig. 2e). The text does not name the condition pair, and the Fig. 2e heatmap prints no values. (Main 'Modeling tissue-specific splicing' paragraph 4; Fig. 2e)
• SpliceAI2 share of excess rare-disease burden: Excess burden share Among excess panel-matched variants, protein-truncating variants accounted for 50% and SpliceAI2-identified cryptic splice variants a further 15%; half of those cryptic splice variants lie more than 50 bp into introns. At panel-match odds ratio 2 the next best model recovered 114 excess variants; the text does not name it. (Main 'Contribution to rare genetic disease' paragraphs 2 and 3; Fig. 4e)
• SpliceAI2 tissue-specific training-strategy and variant-effect checks: Ablation results Extended Data Fig. 4a prints Pearson r 0.37 (one-hot, from scratch), 0.38 (147 RBP, from scratch), 0.43 (one-hot, fine-tune) and 0.44 (147 RBP, fine-tune). Extended Data Fig. 4c prints Pearson r 0.5393 for ridge regression on measured reference usage alone and 0.5395 with SpliceAI2 tissue predictions added; the authors state there was no additional benefit from tissue-specific predictions for tissue-specific variant effects. (Extended Data Fig. 4a and 4c (page 55); Main 'Modeling tissue-specific splicing' paragraph 5)
• SpliceAI2 top-transcript reconstruction: Top transcript reconstruction Perfectly reconstructed the most abundant transcript for 82% of held-out genes (chromosomes 21 and 22, n = 352), compared with 78% when trained without long-read data (human-only models). (Main 'Hierarchical model of RNA splicing' paragraph 2; Methods 'Top transcript benchmark'; Fig. 1c)
• SpliceAI2 training data: Training data Trained on splice site and junction usage from 314,745 short-read RNA-seq samples (67,493,026 junctions) in human and nine other mammalian species, and on transcript usage from 330 ENCODE long-read samples in human and mouse. Human: 127,431 samples from recount3 (GTEx, TCGA, CCLE and other SRA-derived projects). (Main 'Hierarchical model of RNA splicing' paragraph 1; Methods 'Training datasets'; Supplementary Table 1 rows 'Homo_sapiens' and 'Total')
• SpliceAI2 training split and benchmark overlap: Training split Genes on human chromosomes 21 and 22 were held out for validation and all other genes were used for training. The variant benchmarks (GTEx and GEL GT/AG, GTEx sQTL, OpenSplice, UK Biobank) are not restricted to held-out chromosomes, so most benchmark variants lie in genes seen in training; only the top-transcript benchmark is restricted to chromosomes 21 and 22. The paper does not say whether junctions observed in GTEx benchmark samples were excluded from training. (Methods 'SpliceAI2 model' (Training, last paragraph), 'Top transcript benchmark' and 'Variant effect benchmarks'; Supplementary Table 1 row 'Homo_sapiens')

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.