Genomics · global
Reading RNA Splicing Clues from DNA: Illumina Releases SpliceAI2 to Help Identify Rare Disease–Associated Variants
Genetic clues to rare diseases sometimes lie in the details of how genes are spliced into RNA. SpliceAI2 identifies more candidate variants in existing research data, but whether this advance can translate into more confirmed diagnoses still requires clinical validation.
For families seeking the cause of a rare disease, reading the DNA sequence is only the starting point; the challenge often lies in determining which variant truly disrupts gene function. Illumina released SpliceAI2 on October 8 to try to address one part of this gap: predicting how DNA changes affect RNA splicing so researchers have a stronger basis for selecting genetic clues worth investigating.
When cells produce mature RNA, they remove segments called introns and join the retained exons. If splice sites change, the resulting products may also change. According to Illumina’s technical overview, SpliceAI2 can predict splice site usage frequencies, connections between sites, and complete RNA transcripts directly from DNA sequences. This extends variant interpretation from “where splicing might go wrong” to “what might ultimately be produced.”
The company’s claim of identifying “approximately 17% more disease-associated variants” has a specific statistical basis. The research manuscript analyzed 7,504 probands from Genomics England, focusing on diseases in which a single allele can cause disease, and shuffled patient phenotype pairings to estimate the background of chance matches. At the same odds ratio of 2 for matches to disease gene lists, SpliceAI2 identified 133 variants above the number expected by chance, compared with 114 for the next-best model. This is where the approximately 17% difference comes from; it does not mean a 17% increase in the rate of confirmed diagnoses.
RNA data provided another layer of testing: in a subset selected by phenotype and model scores, 17 of 23 rare variants prioritized for follow-up showed the predicted splicing changes. The manuscript also noted that approximately half of the cryptic splicing–associated variants identified were located within introns, more than 50 bases from the boundary, and might typically be missed by exome sequencing. This also highlights that the usefulness of prediction tools depends on whether the original testing covers those locations.
The model’s training data included 314,745 RNA sequencing samples from ten species and more than 46 million splice junctions after filtering. Illumina stated that adding 330 ENCODE long-read samples increased the accuracy of complete reconstruction of the predominant transcript from 78% to 82% in genes not used for training. This test measures transcript reconstruction capability and addresses a different question from the comparison of rare disease candidate variants.
The tool is also available. PyPI records show that SpliceAI2 2.0 was released on October 8; the documentation provides pretrained weights and precomputed results for 4 billion possible single-nucleotide variants within protein-coding gene regions and 150 million insertion or deletion variants observed in populations. The weights and prediction data are available for academic and noncommercial research; commercial use requires contacting Illumina, while local execution requires a CUDA-compatible GPU.
For rare disease research, these predictions can help guide subsequent functional experiments; for drug discovery, they can provide research clues about how variants alter RNA products. However, the core evidence currently comes from a preprint and analyses of existing data, and prospective diagnostic benefit has yet to be demonstrated. Observing a splicing change also does not mean it has been confirmed to cause a patient’s disease. Incorporating the tool into diagnostic workflows will still require combining phenotypic, genetic, and functional evidence to test whether it can increase reliable confirmed diagnoses and improve interpretation efficiency.