Webcast Recap: CI-SpliceAI for Splice Site Prediction and Variant Interpretation in VarSeq

We’re grateful to everyone who tuned in for our April 15th webcast, “CI-SpliceAI for Splice Site Prediction and Variant Interpretation in VarSeq,” with presenter Nathan Fortier.

If you weren’t able to attend live, here’s a quick overview of what was covered: the ClinGen SVI Splicing Subgroup has released evidence-based recommendations for applying splice prediction scores within the ACMG/AMP variant classification framework, and commercial laboratories have long faced a licensing barrier with the gold-standard SpliceAI tool. This webcast walked through both of those topics and introduced CI-SpliceAI as a validated, commercially licensed alternative now integrated directly into VarSeq.

Webcast Highlights

A major focus of the webcast was the ClinGen SVI Splicing Subgroup’s recommendations, which provide the concrete thresholds and criteria mappings that the original 2015 ACMG/AMP guidelines left unspecified. The recommendations center on SpliceAI delta scores and address five criteria:

1. PVS1 is reserved exclusively for variants directly disrupting canonical ±1/±2 splice dinucleotides;
2. PP3 applies when the maximum SpliceAI delta score is 0.2 or above;
3. PS1 can be combined with PP3 when a previously classified variant shares the same predicted splicing impact; The delta position reported by SpliceAI (the predicted distance from a variant to the affected splice site) is what makes it possible to determine whether two variants share the same predicted impact.
4. BP4 applies when the maximum delta score is 0.1 or below;
5. BP7 can be combined with BP4 for synonymous and intronic variants.

The webcast also addressed a practical problem facing commercial clinical laboratories: SpliceAI’s pre-trained model weights and precomputed scores are restricted to non-commercial use. To determine whether an open-source alternative could serve as a clinical replacement, Golden Helix conducted an independent benchmarking study comparing SpliceAI, CI-SpliceAI, and OpenSpliceAI across three independent datasets.

The headline finding was that CI-SpliceAI achieves statistically equivalent performance to SpliceAI across all three benchmarks, with balanced Spearman correlation across all four delta score types and exact splice site match rates above 90%. OpenSpliceAI performed well on splice loss events but showed notably weaker correlation on splice gain events. Based on these results, CI-SpliceAI was identified as the preferred open-source alternative for clinical use.

Finally, the webcast introduced two new CI-SpliceAI annotation tracks available now in VarSeq for all users with a VSClinical license:

• The CI-SpliceAI Variant Track provides over 47 million precomputed scores natively computed for both GRCh37 and GRCh38, avoiding the liftover artifacts present in Illumina-provided SpliceAI scores and supplemented with ClinVar and other curated variants for comprehensive clinical coverage.
• The High Confidence Regions Track solves a critical interpretation problem by marking genomic positions where an absent variant can be confidently treated as scoring below the 0.1 threshold. BP4 can be applied when all of the following are true:
  • The variant is absent from the CI-SpliceAI track;
  • It overlaps a High Confidence Region;
  • It is a SNP, single-base insertion, or deletion of one to four base pairs.

A live demo walked through a hypothetical case of a patient with renal insufficiency, hypertension, and respiratory distress whose exome sequencing came back negative. CI-SpliceAI was used within VarSeq to identify a deeply intronic cryptic splice variant that would otherwise have been missed. Then, VSClinical was used to review delta scores, visualize the predicted cryptic acceptor site, and generate a final clinical report.

Looking ahead, full automation is coming later this year: the VarSeq auto-classifier will automatically evaluate CI-SpliceAI scores and apply all four evidence codes (PP3, PS1, BP4, and BP7) per ClinGen SVI recommendations, and VSClinical will gain a new splice site visualization tool that displays predicted cryptic splice site locations relative to existing exon structure.

Conclusion

If you missed the live session or would like to revisit any part of the discussion, the full webcast recording is available on demand here. For those interested in diving deeper into our benchmarking methodology and results, our preprint is available on bioRxiv here. If you’re interested in exploring how VarSeq and VSClinical can support splice variant interpretation in your clinical workflows, please contact our technical and sales experts to schedule an evaluation or demo.