Folklore

Documentation / Computational Predictors / SpliceAI

SpliceAI Score Interpretation and ACMG Thresholds

SpliceAI reports donor and acceptor gain or loss. Folklore reads the maximum delta score through explicit thresholds: values at or below 0.1 can support defined benign guards, values from 0.2 enter the PP3_splice path with strength modulation, and PVS1 blocks duplicate computational evidence for the same loss-of-function mechanism.

What the Four Scores Represent

SpliceAI is a deep learning model developed by Illumina that predicts whether a genetic variant will disrupt normal mRNA splicing. Splicing is the process by which the cell removes non-coding sections (introns) from the pre-mRNA and joins the coding sections (exons) to produce the final messenger RNA. Variants that disrupt this process can lead to abnormal proteins or complete loss of protein production, even if they do not directly change the amino acid sequence.

The model was published in Cell (Jaganathan et al., 2019) and is widely used in clinical genetics laboratories. It is one of two computational tools in Folklore that directly influences ACMG classification.

Four Delta Scores

SpliceAI produces four scores, each representing a different type of splice disruption. Each score ranges from 0 (no impact) to 1 (certain disruption). The maximum of the four scores is used for classification thresholds.

ScoreNameMeaning
DS_AGAcceptor GainThe variant creates a new splice acceptor site where one did not exist
DS_ALAcceptor LossThe variant destroys an existing splice acceptor site
DS_DGDonor GainThe variant creates a new splice donor site where one did not exist
DS_DLDonor LossThe variant destroys an existing splice donor site

Clinical Score Thresholds

Max Score RangeInterpretation
0.0 to 0.1Can support defined benign splice guards when the consequence is eligible.
Above 0.1 to below 0.2Indeterminate for the formal splice evidence path.
0.2 to below 0.5PP3_splice Supporting when all eligibility guards pass.
0.5 to below 0.8PP3_splice Moderate when all eligibility guards pass.
0.8 to 1.0PP3_splice Strong when all eligibility guards pass.

How Folklore Uses SpliceAI

SpliceAI scores are used in three distinct ways within the classification engine. Each role has a specific threshold and clinical rationale.

ConditionACMG EffectRationale
0.2 to below 0.5PP3_splice SupportingSupporting splice evidence when consequence, PVS1 and mutual-exclusion guards pass.
0.5 to below 0.8PP3_splice ModerateModerate splice evidence when the same eligibility guards pass.
0.8 or higherPP3_splice StrongStrong splice evidence when disease-association and other eligibility guards pass.
Max score below 0.1 or absentGuard for missense BP4BayesDel benign evidence is blocked when a concerning splice signal is present.
Max score <= 0.1BP7 and BP4_splice guardsBenign splice evidence is consequence-specific. Canonical splice-motif and other ineligible consequences remain excluded.

PVS1 Double-Counting Guard

When PVS1 (loss-of-function) is triggered for a variant, PP3_splice is not applied. This prevents counting splice disruption leading to loss of function as both PVS1 and PP3 evidence. Pure missense variants with BayesDel data but no VEP splice consequence do not receive a second computational splice criterion. When both missense and splice consequences are present, the classifier uses mutual-exclusion safeguards so correlated predictions do not become additive votes.

Score Source

SpliceAI scores are precomputed from Ensembl MANE transcript predictions (Release 113). They are not computed at runtime. This ensures reproducibility -- the same variant always receives the same SpliceAI scores regardless of when the analysis is run.

Limitations

SpliceAI predictions are computational. RNA splicing studies (RT-PCR, minigene assays) remain the gold standard for confirming splice-altering effects. Scores are computed on MANE Select transcripts only, so non-MANE transcript-specific splicing effects may be missed. Deep intronic variants beyond the SpliceAI prediction window (typically +/- 50bp from precomputed scores) may not be captured. The model was trained on known splice sites, so novel splice mechanisms not represented in the training data may not be predicted.

Clinical Relevance

Approximately 10-15% of pathogenic variants cause disease through aberrant splicing. SpliceAI provides a fast, reproducible screen for splice impact that can guide whether RNA studies are warranted. When reviewing variants with high SpliceAI scores, consider recommending RT-PCR or minigene assay confirmation.

Reference: Jaganathan K, et al. Cell. 2019;176(3):535-548. PMID: 30661751

ClinGen SVI: Walker LC, et al. Am J Hum Genet. 2023;110(7):1046-1067. PMID: 37352859