On October 8, Illumina released SpliceAI2, a deep learning model that predicts how DNA variants change RNA splicing, and on October 14 it launched a software platform called BioInsight that packages it. Illumina's release says SpliceAI2 identified 17% more disease-relevant variants than other splicing models in a rare disease dataset.

That claim is about one kind of variant. To see how much of the interpretation problem it touches, we counted what ClinVar, the public archive of clinical variant classifications, still labels uncertain.

How big is the uncertain pile?

We queried NCBI's ClinVar through E-utilities for records by molecular consequence and by clinical significance.

Source: NCBI ClinVar via E-utilities, queried October 19, 2026. Counts are ClinVar records. A record can carry more than one consequence term, so classes overlap. Percentages are VUS records divided by all records with that consequence term. Total ClinVar records at query time: 4,687,256, of which 2,419,589 are VUS (51.6%).

View data

class	value
Missense	84.1
Intron	25.7
Nonsense	12.5
Frameshift	11.9
Canonical splice site	11.2
Synonymous	3.4

ClinVar held 4,687,256 records, and 2,419,589 of them (51.6%) were classified as uncertain. By contrast only 447,149 were pathogenic or likely pathogenic, so there were 5.4 uncertain records for each one.

The uncertainty is concentrated in missense variants. They account for 2,175,987 uncertain records, or about 90% of all of them. Records where the change disables the protein outright are rarely uncertain: 12% of nonsense and 12% of frameshift records. Canonical splice-site variants, the two bases at each end of an intron, are similar at 11.2%, and 52% of them are pathogenic or likely pathogenic.

Where does a splicing model help?

Intron variants are the middle case. ClinVar lists 1,036,181 records with an intron consequence. Of those, 266,055 (25.7%) are uncertain and 44,319 (4.3%) are pathogenic or likely pathogenic, a ratio of 6.0 to 1. Those 266,055 records are about 11% of all uncertain records.

That is the space a splicing predictor addresses: variants near or deep inside introns whose effect on an exon is hard to read from sequence alone. In Illumina's preprint, predicted splice-altering variants account for 15% of the excess genetic burden in a Genomics England rare disease cohort, and about half of those sit in deep intronic regions that exome sequencing is likely to miss.

SpliceAI2 is not limited to introns; missense variants can also alter splicing. We did not test how many of ClinVar's uncertain records SpliceAI2 would reclassify.

What exactly is claimed?

The preprint is posted by Illumina, with authors from Illumina and Oxford, and describes a model trained on 314,745 RNA sequencing samples from human and nine other mammalian species plus 330 long-read samples. The press release says it improved splice site usage quantification by 34% over the next best model. We did not find the 17% comparison baseline stated in the preprint text we read, so we treat it as a company claim. The work is not peer-reviewed, and the model is available through Illumina's BioInsight applications, not as an open release that we could verify.

What this does and does not show

The ClinVar counts show where classification effort has not caught up, not where the model succeeds. A variant leaves the uncertain pile only after evidence is assembled by a laboratory or an expert panel, and a prediction is one piece of that evidence. We have not checked SpliceAI2 against ClinVar. ClinVar is also a submission archive, not a random sample of the genome, and its mix reflects what clinical labs have tested.

Sources

1. NCBI, ClinVar, E-utilities queries on October 19, 2026 by molecular consequence and by clinical significance property; every percentage and ratio above was recomputed in code.
2. Illumina, Illumina releases SpliceAI2 to help advance rare disease research, October 8, 2026: the 17% and 34% claims.
3. Jaganathan et al., A unified framework for quantitative splicing and transcript prediction (SpliceAI2 preprint): training data, the 15% excess-burden result and the deep intronic share.
4. Illumina, BioInsight Platform launch, October 14, 2026.
5. Our earlier analysis: 9 billion predictions, 4.6 million records: what AlphaGenome Atlas can and cannot fill in.