Model type
Dilated convolutional splicing predictor
Configuration
SpliceAI 1.3.1 · mask 0 (S0)
SpliceAI annotates sequence variants with predicted splice acceptor and donor changes.
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and LicenseHow it worksSpliceAI workflow
Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation.
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and LicenseOverview
Inputs
VCF variants, reference FASTA and matching gene annotation, or custom one-hot-encoded sequence.
Outputs
Acceptor/donor gain/loss scores and positions in VCF INFO annotations.
Access
Official project documentation and implementation: https://github.com/illumina/SpliceAI
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and Licenselimited source coverage · Automated source review, 2026-09-16. All specifications and missing details
Evaluations and results
1 evaluation · 4 results. Different protocols are not a single leaderboard.
Filter evaluations
Applied filters: All linked evaluations
Source checking is not independent reproduction. Release 2026-09-25-d40cee0abe73.
Use this model
How it works, versions and access
Related profile: SpliceAI. This page retains the exact record and its evaluation context.
This configuration
8,297 of 8,324 held-out variants scored in every configuration (314 positives, 460 exon groups). The same 27 rows are excluded: 23 assembly-orientation mismatches and four outside the selected canonical transcript spans. Missing scores are not zero or negative predictions.
- record
- SpliceAI 1.3.1 · mask 0 (S0)
- configuration
- Not reported
- entity type
- Configuration
How it works
How it works
SpliceAI reads one-hot-encoded DNA through dilated convolutional residual blocks. Skip connections combine features at different depths, and a softmax layer assigns acceptor, donor or neither probabilities to the central positions. The 10kb version requires 5kb of sequence on each side of a scored position; variant scoring compares the reference and alternate predictions.
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and LicenseVersions and reproducibility
The paper studies 80nt, 400nt, 2kb and 10kb receptive spans. Variant scoring averages five independently trained models; these are not five different assay results. SpliceAI-10k uses 5,000 flanking bases on each side. An input of length l + 10,000 produces predictions for l central positions; receptive span is distinct from maximum input length.
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and LicenseStrengths, limitations and unresolved questions
Strengths and limitations
Strengths and considerations
- Provides direct sequence inference and an annotation workflow with explicit genome and distance settings.
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License
Limitations and conditions
- The command-line pipeline skips unsupported variants and variants outside its gene annotations. Code, model weights and downloadable precomputed scores have distinct licensing provisions.
Sources (2)
illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License
Profile review details
Inspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied.
Stable record: discovery-model-spliceaiSpecifications
Inputs, training, access and other details
Explanatory profile: limited source coverage · Automated source review, 2026-09-16. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.
| Property | Description and evidence |
|---|---|
| Model type | Dilated convolutional splicing predictorSources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Architecture | Residual one-dimensional convolutional network with dilated kernels and skip connections; a softmax head predicts acceptor, donor and neither at each central position.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Inputs | VCF variants, reference FASTA and matching gene annotation, or custom one-hot-encoded sequence.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Outputs | Acceptor/donor gain/loss scores and positions in VCF INFO annotations.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Parameters | The original STAR Methods specifies residual blocks, dilation and receptive spans, but does not state a complete parameter count for each released five-model scoring ensemble. · Not reported in inspected sourcesSources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Known versions | The paper studies 80nt, 400nt, 2kb and 10kb receptive spans. Variant scoring averages five independently trained models; these are not five different assay results.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Training data | Human GRCh37 sequence and GENCODE V24lift37 principal protein-coding transcripts, split by chromosome with non-paralogous held-out test genes. The paper distinguishes GENCODE-only training from GTEx-junction-augmented models used for variant analyses.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Training cutoff | GENCODE V24lift37 on GRCh37 defines the documented transcript annotations. GTEx-augmented training is separately described; the paper does not give one common latest-data date for both variants.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Context limits | SpliceAI-10k uses 5,000 flanking bases on each side. An input of length l + 10,000 produces predictions for l central positions; receptive span is distinct from maximum input length.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Weights licence | CC-BY-NC-4.0 for trained models; commercial use requires a separate licence. Code is PolyForm Strict 1.0.0.Sources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Access | Official project documentation and implementation: https://github.com/illumina/SpliceAISources (2)illumina/SpliceAI: README.md; spliceai: Primary paper PDF · Jaganathan et al., Cell 2019, STAR Methods: SpliceAI architecture and Model training and testing (pp. e2–e3); README.md: usage and License |
| Code licence | PolyForm Strict 1.0.0 for code; trained weights have separate terms.Sourcesillumina/SpliceAI: LICENSE · LICENSE: licence text |
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
5 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Relationship: variant of discovery-model-spliceai Individual claims | MFASS matched canonical annotation v1: exclusion-verification.json manifest-v1.json: study and conditions Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 4be7a98e2553fa2378c29625b13eb3e8ac2e58fb | source checked automated execution evidence review · 2026-09-25T11:24:56Z Audit detailsComputational source, coverage and metric checks only; not independent human review or published-score reproduction. Field: Claim: rewire-mfass-matched-v1-configuration-s0-variant-of-discovery-model-spliceai Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Relationship: variant of discovery-model-spliceai Individual claims | MFASS matched canonical annotation v1: manifest-v1.json manifest-v1.json: study and conditions Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 093fd1ae198c80ce34408d84d6543bca4fc538f2 | source checked automated execution evidence review · 2026-09-25T11:24:56Z Audit detailsComputational source, coverage and metric checks only; not independent human review or published-score reproduction. Field: Claim: rewire-mfass-matched-v1-configuration-s0-variant-of-discovery-model-spliceai Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Relationship: variant of discovery-model-spliceai Individual claims | MFASS matched canonical annotation v1: provenance.json manifest-v1.json: study and conditions Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 093fd1ae198c80ce34408d84d6543bca4fc538f2 | source checked automated execution evidence review · 2026-09-25T11:24:56Z Audit detailsComputational source, coverage and metric checks only; not independent human review or published-score reproduction. Field: Claim: rewire-mfass-matched-v1-configuration-s0-variant-of-discovery-model-spliceai Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Relationship: variant of discovery-model-spliceai Individual claims | MFASS matched canonical annotation v1: report.json manifest-v1.json: study and conditions Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 093fd1ae198c80ce34408d84d6543bca4fc538f2 | source checked automated execution evidence review · 2026-09-25T11:24:56Z Audit detailsComputational source, coverage and metric checks only; not independent human review or published-score reproduction. Field: Claim: rewire-mfass-matched-v1-configuration-s0-variant-of-discovery-model-spliceai Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Relationship: variant of discovery-model-spliceai Individual claims | MFASS matched canonical annotation v1: verification.json manifest-v1.json: study and conditions Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 093fd1ae198c80ce34408d84d6543bca4fc538f2 | source checked automated execution evidence review · 2026-09-25T11:24:56Z Audit detailsComputational source, coverage and metric checks only; not independent human review or published-score reproduction. Field: Claim: rewire-mfass-matched-v1-configuration-s0-variant-of-discovery-model-spliceai Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Sources and history
View linked audit checks and correction history
Release 2026-09-25-d40cee0abe73 · Record review: source checked
5 source records and release history
- MFASS matched canonical annotation v1: report.json · Original source · 093fd1ae198c80ce34408d84d6543bca4fc538f2
- MFASS matched canonical annotation v1: manifest-v1.json · Original source · 093fd1ae198c80ce34408d84d6543bca4fc538f2
- MFASS matched canonical annotation v1: verification.json · Original source · 093fd1ae198c80ce34408d84d6543bca4fc538f2
- MFASS matched canonical annotation v1: provenance.json · Original source · 093fd1ae198c80ce34408d84d6543bca4fc538f2
- MFASS matched canonical annotation v1: exclusion-verification.json · Original source · 4be7a98e2553fa2378c29625b13eb3e8ac2e58fb
Technical metadata and extraction receipts
Stable ID: rewire-mfass-matched-v1-configuration-s0
- areas
- dna-genomes
- condition
- S0
- configuration
- tool: spliceai; mask: 0; stem: spliceai-1.3.1-gencode44-canonical-mask0
- distance
- 50
- annotation
- GENCODE v44 canonical transcripts
- code
- spliceai version: 1.3.1; spliceai files: init .py: sha256: 5e0ade84087ee01040dbc3344359cbfdb26e2fcf98c214e15110ac1d05562037; git blob: b5d843e672cfff52c6d050205dead13d7c31533e; utils.py: sha256: 207e147c73bdc1970fdb176dcbf088883f9a58756bfeb0d93494e75018ae477c; git blob: 66f2c3c14198d9020a7913f0412fc91ade008b91; main .py: sha256: 1090e97a5f7d8f937048b7d1f08d98fb593bd86953e9d64178aa3ca096961a77; git blob: 2c59acd61ce528dd0ed6d6c14244579354fed146; models/spliceai1.h5: sha256: e1fd5adcef7489d604b10e79c40078ef790d51ef048c4ce3869c9119ac5de42b; git blob: e5e6bf5d2f61f9f9f8f1c93386881992a96a82c0; models/spliceai2.h5: sha256: 6ab042b82ab966b6d3582cb31b96f0859ea08a864f168d69e83aa14450a3b66e; git blob: b6c197b52fc39e81f42a2dfb25eeef11db3c94af; models/spliceai3.h5: sha256: e2e790bde53dfdf410c6dc434a86122a7d12f3f38dc2ef45d85986e9ecf22fad; git blob: 5009694c892a782faa70f66e23f0f74cf69c6809; models/spliceai4.h5: sha256: ca88ac9e58e69ba6fdeed319b72f063f164c9abf7392eaccef903e94c1d99dd6; git blob: 81ce1c887855f34abf9af5f5c65d85fd04d3132d; models/spliceai5.h5: sha256: 791cd22c62a80a08d2ca674615a93ce8159d7b55bd157cfef2983b1bd6b41391; git blob: 752679198e4f52f69b93945b2c49e4efc8534860; spliceai wheel only files sha256: clean main .py: d91e273a60b1b1c079cd29974f1c9901d997bc6f57226ff094ad15d0c1e49b03; normalise chrom.py: 01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b; spliceai source identity: upstream-b3c7f17; spliceai upstream files verified: true; spliceai upstream file mismatches: None recorded
- resources sha256
- cohort: 389702ff4c647d7ce10a90092a6fa811ae777d15997baf39ce9aae0346247bd0; split: 999ebcb7e63a5c5eaa8780fa468e59ac1f934260ad50102814174c396317f052; fasta: e49b92b3e4f321bf254c042f25b726d9931c4d74c7523e8b6bb530e63b0cfd4b; fai: a2c323ea4cff34d7123ace4578f7e122b2d2f5a22f40dc23eb8b97d17723d169; spliceai annotation: 2383cd148b292b9a6af56da160c9c142ddeb5720b1e9ef7c5966a1744f684c4d; pangolin db: 84ae1a18d9d7a32e14e3fe9c8705256a28cd9bde0a55bc46200cbe77178deb27
- source locator
- manifest-v1.json: conditions.S0; code.spliceai
- limitations
- 8,297 of 8,324 held-out variants scored in every configuration (314 positives, 460 exon groups). The same 27 rows are excluded: 23 assembly-orientation mismatches and four outside the selected canonical transcript spans. Missing scores are not zero or negative predictions.; Exploratory comparison: prior results were known. Paired contrast intervals are unadjusted and do not establish a universal model ranking.; Pangolin uses the recorded per-gene masking patch; these are exact configurations, not unqualified upstream model scores.; Precision at 100 is sensitive to tied-score ordering, especially P1. Numerical source checking is automated, not human review or independent reproduction.; Assembly-orientation issue reported at https://github.com/KosuriLab/MFASS/issues/1. Original v1 outputs remain unchanged; corrections require a new version.