AI and RNA splicing: from prediction to validation

SpliceAI (ClinGen Biocurator Working Group)

ClinGen Resource · 2020-07-24 · Source

English-language educational presentation recorded by ClinGen on July 23, 2020 and published on July 24. Kishore Jaganathan explains the original SpliceAI model; this is not coverage of SpliceAI2 or a new clinical authorization.

Research: what information comes from the sequence?

RNA splicing removes segments from precursor RNA and joins the remaining pieces. AI models investigate how sequence changes might affect this process. The original SpliceAI developers use long surrounding sequence context to predict potential splice sites. Predictions for the reference and variant-containing sequences are compared. This output is not RNA directly measured in an experiment.

Why does tissue matter?

The 2022 Pangolin study examines splicing predictions for heart, liver, brain and testis. The model predicts splice-site usage as well as whether a site is present. The researchers note that capturing differences between tissues remains difficult. Performance on the studied tasks does not imply equal reliability in every tissue or for every variant.

Recent development: the October 8, 2026 SpliceAI2 announcement

Illumina announced SpliceAI2 for rare disease research. The company reports identifying 17% more disease-relevant variants than other splicing models in a research dataset. This is a manufacturer-reported result, not independent clinical acceptance or evidence that 17% more patients received diagnoses. The original SpliceAI and Pangolin publications concern different versions and tasks; their figures are not directly interchangeable.

Clinical interpretation: RNA experiments also need context

ClinGen’s clarification, updated March 22, 2024, addresses how to weigh experimental effects when multiple transcripts are observed. Interpreting incomplete splicing may require tissue-specific and gene–disease-specific assessment. One AI score or one RNA result cannot replace the full clinical evidence.

Separate regulatory authorization from research evidence

This article concerns research models and interpretation of evidence. No publication, expert-group clarification or manufacturer announcement is presented here as a regulatory authorization document for a product. Any proposed clinical use requires separate consideration of intended use and the relevant product’s applicable documentation. This is not individual genetic interpretation or treatment advice.

Editorial proposal: keep a review record alongside the score

Our responsible-AI recommendation is to record the model version, reference genome, evaluated transcript and chosen thresholds together. Show the predicted effect, experimentally observed effect and expert interpretation separately. When they disagree, keep the uncertainty visible instead of producing a single “correct” label. These are editorial recommendations, not measured outcomes of the studies above or claims about the effectiveness of RAI products.