Illumina's new SpliceAI2 finds 17% more disease-linked gene variants other splicing models miss

Illumina said on Thursday, Oct. 8, 2026, that it is releasing SpliceAI2, a new genomic AI model meant to catch disease-relevant splice variants that other tools miss. On a rare-disease research dataset, Illumina says SpliceAI2 identified 17% more disease-relevant variants than other splicing models, and that together with its PromoterAI and PrimateAI-3D models the suite can now flag up to twice as many variants with predicted biological impact.

Most of the genome is not the tidy protein-coding bits textbooks drew in chalk. A huge share of rare-disease mystery lives in how a gene gets cut and pasted into RNA, including splice sites buried deep in introns that look like junk until they aren't. Illumina's first SpliceAI already became the default tool researchers reach for; shipping a bigger, broader sequel and putting the weights on GitHub is the company betting that better AI annotation, not just more sequencing machines, is how you turn a VCF file into an answer a family can act on. The 17% lift and the "2x with the suite" claim are Illumina's own benchmarks, so treat them as the seller's scorecard — still, if even half of that holds up in other labs' hands, that is a lot of previously invisible candidates suddenly on the review list. The catch stays the same as every variant predictor: a confident score is not a diagnosis, and a prettier model can make a wrong call feel more scientific. But for rare disease, hereditary cancer research, and drug discovery, missing the splice is how cases stay unsolved. This is Illumina trying to make fewer of those misses.

On Thursday, 8 October 2026, Illumina introduced SpliceAI2, a genomic AI model for predicting the functional consequences of genetic variation that affects RNA splicing. The company release, carried on PR Newswire, is titled “Illumina releases SpliceAI2 to help advance rare disease research.” The wire stamps 8 October 2026, 9:15 a.m. Eastern, which is 13:15 UTC. Those lines are Illumina’s.

Splicing is the step that cuts and pastes RNA transcripts before a cell makes a protein. A transcript is the RNA copy of a gene. Cryptic splice sites are hidden cut-points, often deep in the long stretches between the protein-coding bits, that can scramble that assembly. Illumina says those hidden sites add a significant share of rare-disease cases and are hard to catch with older methods. Those lines are Illumina’s.

On a rare-disease research dataset, Illumina says SpliceAI2 identified 17% more disease-relevant variants than other splicing models. A variant is a spelling change in DNA. Disease-relevant, here, means a change the company says is more likely to matter for a rare-disease case. That 17% figure is Illumina’s.

Illumina says SpliceAI2, PromoterAI, and PrimateAI-3D together let researchers flag up to twice as many variants with predicted biological impact. PromoterAI looks at the on-switch region of a gene. PrimateAI-3D looks at missense variants, spelling changes that swap one protein building block for another. That “up to twice as many” line is Illumina’s.

A BioInsight preprint compared SpliceAI2 with the original SpliceAI and other splicing models. Illumina says analyses of Genomics England data found 17% more disease-relevant splice variants; analysis of NIH GTEx data found 34% better quantification of how often a splice site is used, versus the next-best model; and the scores lined up more closely with protein-level effects in UK Biobank. GTEx is a public atlas of how genes are used in different human tissues. Those figures are Illumina’s.

The original SpliceAI, released in 2019, is cited in more than 3,400 publications and sits in ClinGen’s splice-variant interpretation guidelines. ClinGen is a clinical-genomics group that writes those standards. Illumina says SpliceAI2 was trained on a dataset 100 times larger than the first model. Those lines are Illumina’s.

The BioInsight article says SpliceAI2 predicts splice sites, the junctions that connect them, and complete transcript isoforms — the finished RNA versions of a gene — from DNA sequence alone. Training included 314,745 RNA-seq samples across 10 species and 330 long-read samples from ENCODE. RNA-seq is a count of which RNA pieces a sample actually made. Long-read sequencing reads a transcript end to end instead of in scraps. Those counts are Illumina’s.

Researchers can use SpliceAI2 through Illumina’s DRAGEN Annotation and Emedgene products. The BioInsight article also points to GitHub, at github.com/Illumina/SpliceAI2, for source code, trained models, and precomputed predictions for every possible single-letter DNA change inside human gene bodies, plus insertion-and-deletion changes seen in human populations. Those lines are Illumina’s.

Rami Mehio, senior vice president and general manager of BioInsight, said: “Variant effect prediction tools, such as SpliceAI2, are among the key areas of focus for the BioInsight AI Lab. As researchers work to elucidate the effect of mutations, we are uniquely positioned to unite genomic data and scientific expertise at scale, delivering the AI tools that can advance discovery and human health.” That statement is on Illumina’s release.

Kyle Farh, vice president of Illumina’s BioInsight AI Lab, said: “Illumina is advancing AI to systematically shrink the portion of the genome that remains uninterpretable. Genomics has driven some of the most consequential genetic disease breakthroughs of the past two decades. Today, we are equipping researchers with the next generation of technology to help understand the underlying causes of disease.” That statement is on Illumina’s release.

The picture is Illumina’s official SpliceAI2 rare-disease research release graphic: a gold DNA helix on a pale orange field, with the company’s announcement line. It is the company’s launch image. The frame does not print a calendar date.

In plain terms, Illumina said on Thursday that it is releasing SpliceAI2, an AI model that predicts how a DNA change will scramble the way a gene is cut and pasted into RNA. On its rare-disease research set, the company says the model found 17% more disease-relevant variants than other splicing tools, and that with PromoterAI and PrimateAI-3D the suite can flag up to twice as many variants with a predicted biological effect. The original 2019 SpliceAI is already a standard citation. The new weights and code are on GitHub. The 17% and “up to 2x” figures are Illumina’s own scorecard.