Documentation / Reference Databases / SpliceAI Precomputed
SpliceAI Precomputed
SpliceAI is a deep learning model developed by Illumina that predicts the impact of genetic variants on mRNA splicing. Helix Insight uses precomputed SpliceAI scores for all coding variants, eliminating the need for on-the-fly prediction and ensuring consistent, reproducible results.
Database Details
Four Delta Scores
SpliceAI produces four delta scores, each ranging from 0 to 1, representing the change in splice probability caused by the variant:
Predicts creation of a new splice acceptor site. A high score indicates the variant may introduce a cryptic acceptor competing with the canonical site.
Predicts destruction of an existing splice acceptor site. A high score indicates the canonical acceptor is disrupted.
Predicts creation of a new splice donor site. A high score indicates a cryptic donor may be introduced.
Predicts destruction of an existing splice donor site. A high score indicates the canonical donor is disrupted.
The maximum of the four delta scores (max_score) is used for classification thresholds. This captures the strongest predicted splice effect regardless of mechanism.
Score Interpretation
| Max Score | Interpretation | Clinical Implication |
|---|---|---|
| 0.0 - 0.1 | No predicted splice impact | Variant unlikely to affect splicing. BP7 guard satisfied (SpliceAI < 0.1). |
| 0.1 - 0.2 | Low predicted impact | Some splice effect possible but below PP3_splice threshold. |
| 0.2 - 0.5 | Moderate predicted impact | PP3_splice triggered (Supporting). ClinGen SVI 2023 threshold for supporting splice evidence. |
| 0.5 - 0.8 | High predicted impact | Strong prediction of splice disruption. PP3_splice at Supporting strength. |
| 0.8 - 1.0 | Very high predicted impact | Near-certain splice disruption. PP3_splice at Supporting strength. |
Role in ACMG Classification
SpliceAI max_score >= 0.2 triggers PP3 at Supporting strength. Applies only when PVS1 does not already apply (ClinGen SVI double-counting guard). Aligned with Walker et al. 2023 ClinGen SVI recommendations.
BP7 (synonymous + no splice impact) requires SpliceAI max_score < 0.1 to confirm the variant does not affect splicing. Without this guard, a synonymous variant near a splice site could be incorrectly classified as benign.
BP4 (computational evidence suggests no impact) additionally requires SpliceAI max_score < 0.1 to ensure no predicted splice disruption before applying benign computational evidence.
PVS1 Double-Counting Guard
When a variant triggers PVS1 (null variant in a LoF-intolerant gene) through a splice consequence (splice_acceptor_variant or splice_donor_variant), SpliceAI PP3_splice is not additionally applied. This prevents counting the same splice disruption evidence twice -- once through the consequence-based PVS1 pathway and again through the SpliceAI prediction pathway. This follows ClinGen SVI 2023 guidelines.
Limitations
SpliceAI predictions are based on primary sequence context. Tissue-specific splicing regulation is not modeled.
Precomputed scores cover coding variants in MANE Select transcripts. Variants in non-MANE transcripts or deep intronic regions may not have scores.
SpliceAI does not predict the functional consequence of aberrant splicing (exon skipping, intron retention, etc.), only the probability that splicing is disrupted.
Scores near the 0.2 threshold should be interpreted with caution. RNA studies can confirm or refute predicted splice effects.
References
Jaganathan K, et al. "Predicting splicing from primary sequence with deep learning." Cell. 2019;176(3):535-548. PMID: 30661751.
Walker LC, et al. "Using the ACMG/AMP framework to capture evidence related to predicted and observed impact on splicing." American Journal of Human Genetics. 2023;110(7):1046-1067. PMID: 37352859.
For more details on SpliceAI interpretation, see the dedicated SpliceAI predictor page.