Illumina's New AI Model Predicts Hidden Genetic Disease Culprits Without Expensive RNA Testing

Illumina's BioInsight AI Lab has unveiled SpliceAI2, an artificial intelligence model that predicts how genetic variants affect RNA splicing and gene transcript formation directly from DNA sequences—eliminating the need for costly, tissue-specific RNA sequencing that has traditionally been required for comprehensive splicing analysis. The advancement addresses a critical blind spot in genetic disease diagnosis: over 95% of human genes undergo alternative splicing, yet many disease-causing variants affecting this process remain undetected by conventional genetic testing focused on protein-coding regions.

The model represents a substantial leap beyond the original SpliceAI, which has been cited in more than 3,400 publications since its 2019 release. Trained on a dataset 100 times larger than its predecessor—encompassing 314,745 RNA sequencing samples across ten species, 46 million observed splice junctions, and 330 long-read samples from the public ENCODE project—SpliceAI2 demonstrated measurably superior performance in clinical validation studies.

Clinical Validation Shows Diagnostic Gains

When tested on data from 7,504 individuals in Genomics England's 100,000 Genomes Project, SpliceAI2 identified 17% more disease-associated variants than competing splicing models at matched confidence thresholds. The model reconstructs the most common transcript isoform 82% of the time for previously unseen genes, compared to 78% accuracy when using short-read RNA sequencing data alone.

This improvement carries significant implications for rare disease diagnosis, where causative variants frequently lie in non-coding regions that affect splicing but are overlooked by traditional analysis. For patients who have undergone inconclusive genetic testing, SpliceAI2's enhanced ability to interpret variants of uncertain significance could provide long-sought answers.

Outperforming the Competition

Independent benchmarking across three datasets showed SpliceAI2 outperformed the original SpliceAI, Pangolin, and Google DeepMind's AlphaGenome models. Researchers at the University of Oxford conducted independent comparisons with AlphaGenome, validating Illumina's performance claims.

The model's capabilities extend beyond identifying splice sites to predicting complete transcript isoforms—a critical distinction given that alternative splicing can produce proteins with opposite functions from the same gene. The BCL2L1 gene exemplifies this complexity: it produces both a long isoform (BCL-xL) that prevents cell death and a short isoform (BCL-xS) that promotes it, with the balance between them influencing cancer progression.

Implications for Diagnostics and Drug Development

For genetic testing laboratories, SpliceAI2 offers the potential to enhance variant interpretation pipelines without requiring expensive long-read RNA sequencing infrastructure. The ability to predict splicing outcomes from DNA alone could reduce testing costs while expanding analysis beyond traditional protein-coding sequences.

Pharmaceutical companies may find new therapeutic opportunities in the model's tissue-specific splicing predictions, particularly for cancers driven by aberrant splicing patterns. Understanding how genetic variations affect transcript outcomes could accelerate development of splice-modifying drugs targeting diseases where isoform imbalance contributes to pathology.

For healthcare providers and genetic counselors, the technology promises more comprehensive variant interpretation without the logistical challenges of obtaining difficult-to-access tissue samples for RNA analysis—a particularly valuable capability when investigating rare diseases affecting specific organs or developmental stages.

Sources

1. Introducing SpliceAI2: The Next Generation of Splicing and Transcript Isoform Prediction — Illumina