The Pulse
Illumina’s SpliceAI2 Finds 17% More Rare-Disease Variants
Illumina says its SpliceAI2 model identified 17% more disease-relevant splice variants than other tested splicing models in a rare-disease research dataset. A preprint reports the comparison used 7,504 probands and found 133 excess panel-ma
AI.info Team · Written with AI ·
Illumina says its new genomic AI model found 17% more disease-associated splice variants than other tested prediction models in an analysis of rare-disease cases. Announced October 8, 2026, SpliceAI2 predicts how DNA changes affect RNA splicing—the process cells use to assemble RNA transcripts before making proteins. The results come from a preprint by Illumina’s BioInsight AI Lab, and the company labels the model for research use only, not diagnostic procedures.
What the 17% gain measures
The researchers analyzed whole-genome and phenotype data from 7,504 probands in the Genomics England 100,000 Genomes Project. They assigned disease-associated genes based on participants’ phenotypes and shuffled those phenotype assignments to estimate the background rate of variants matching disease-associated genes. The researchers compared the odds ratio of that match with the number of excess matching variants across different score thresholds. At a fixed odds ratio of 2, SpliceAI2 recovered 133 excess variants, compared with 114 for the next-best model. Those figures amount to roughly 17% more for SpliceAI2, consistent with the percentage Illumina reported.
That comparison does not mean 17% more patients received a diagnosis. The paper estimates that cryptic splice variants identified by SpliceAI2 account for an additional 15% of excess genetic burden beyond protein-truncating variants in its analysis. About half of the prioritized cryptic splice variants lay more than 50 bases into introns, locations that exome sequencing would typically miss.
“Application to a rare disease cohort reveals that predicted splice-altering variants are highly enriched in disease-associated genes and account for 15% of the excess genetic burden, with approximately half residing in deep intronic regions likely missed by exome sequencing.”
Kishore Jaganathan, Kyle Kai-How Farh and their coauthors write in the preprint.
SpliceAI2 predicts more than splice sites
The first SpliceAI, released in 2019, predicts whether a position in a gene is likely to function as a splice site. SpliceAI2 also predicts how splice sites connect and which full-length RNA transcript forms those connections produce. It takes DNA sequence as input, allowing researchers to study potential transcript changes without first obtaining RNA from the tissue where a gene is active.
The model was trained on splice-site and splice-junction measurements from 314,745 RNA-sequencing samples across humans and nine other mammalian species. For transcript prediction, the team also used 330 long-read RNA-sequencing samples. On held-out genes, SpliceAI2 reconstructed the most abundant transcript correctly 82% of the time, compared with 78% when trained without the long-read data.
The preprint also reports comparisons with SpliceAI, Pangolin and AlphaGenome across several variant-effect benchmarks, including cryptic splice-site detection, quantitative trait loci and reporter assays. In one benchmark based on RNA-sequencing data, the model outperformed the other tested models in predicting activation of cryptic splice sites by rare variants. The authors also analyzed population-genomics data and protein measurements, reporting that predicted splice disruption was associated with lower protein abundance among carriers of rare splice-region variants.
Research access, with limits
Illumina says SpliceAI2 is available through its BioInsight applications, including DRAGEN Annotation and Emedgene. The company’s research article also says source code, trained models and precomputed variant predictions are available on GitHub. The original SpliceAI has been cited in more than 3,400 publications, according to Illumina.
The preprint identifies a remaining challenge: predicting how a variant’s effect changes across tissues, cell types and developmental contexts. The paper reports that differences in baseline splicing between tissues were predictive of tissue-specific variant effects, while tissue-specific predictions did not add further benefit in that analysis.