Illumina releases SpliceAI2, a genomic AI model that finds more disease-causing splice variants

Overview

Updated Oct 25

Illumina released SpliceAI2 on October 8. The model predicts how DNA variants disrupt RNA splicing, the process that stitches gene pieces into instructions for proteins.

In Genomics England rare disease data it found 17% more disease-relevant splice variants than other tools. A bioRxiv methods paper posted October 11 said those variants make up about 15% of the cohort's extra disease-linked signal, with about half in deep introns that standard exome sequencing skips. Illumina posted the source code and scores for 4 billion single-nucleotide variants and 150 million indels, and researchers run it on DNA alone, without RNA from hard-to-reach tissues.

Why it matters

Half the extra variants sit in deep introns that standard exome tests skip, so better splice prediction could diagnose patients whose tests came back negative.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

17%

More disease-relevant variants identified

SpliceAI2 found 17% more disease-relevant splice variants than other models in Genomics England data.

34%

Improved splice site usage quantification

SpliceAI2 improved quantification of splice site usage by 34% versus the next best model on GTEx data.

3,400+

Publications citing original SpliceAI

The first-generation SpliceAI has been cited in over 3,400 publications since 2019.

314,745

RNA sequencing samples in training data

SpliceAI2 was trained on RNA sequencing samples from humans and nine other mammalian species.

4 billion

Precomputed single nucleotide variant predictions

Precomputed predictions cover all possible SNVs in human gene bodies, plus 150 million indels.

15%

Excess disease-linked signal from splice variants

Predicted splice-altering variants accounted for about 15% of the extra disease-linked genetic signal in a rare disease cohort, per the bioRxiv preprint.

~50%

Cryptic variants in deep introns

About half of the cryptic splice variants SpliceAI2 found sat more than 50 base pairs into introns, where exome sequencing typically misses them.

Voices

Curated perspectives — historical figures and your fellow readers.

Organizations Involved

Illumina

Genomics company

Developer of SpliceAI2

San Diego-based genomics company that develops sequencing and analysis tools.

Illumina BioInsight Platform

Software platform

Launched in October 2026 as Illumina's unified software environment; includes SpliceAI2

Illumina's unified software environment for multiomics analysis, launched in October 2026, that includes SpliceAI2.

ClinGen

Clinical research consortium

Standards body that incorporated original SpliceAI

Clinical Genome Resource, a consortium that sets standards for clinical genomics.

Genomics England

Government-owned genomics company

Validation partner

UK government-owned company that runs the 100,000 Genomes Project.

Google DeepMind

AI company

Competitor with AlphaGenome

AI research lab that developed AlphaGenome, a competing genomic AI model.

Timeline

2019 October 2026

6 events Latest: October 11th, 2026 · 1 month ago

Tap a bar to jump to that date

October 2026
SpliceAI2 methods preprint posted to bioRxiv
Latest Research

Illumina and academic collaborators posted the SpliceAI2 methods paper to bioRxiv. It reported that predicted splice-altering variants accounted for about 15% of excess genetic burden in a rare disease cohort, with roughly half in deep introns that exome sequencing misses.
October 11th, 2026
Illumina releases SpliceAI2
Product launch

Illumina introduced SpliceAI2, which identified 17% more disease-relevant splice variants than other models in Genomics England data.
October 8th, 2026
SpliceAI2 code and precomputed predictions open-sourced
Product

Illumina published SpliceAI2's source code, trained models, and precomputed predictions on PyPI, GitHub, and Hugging Face. The precomputed set covers all 4 billion possible single-nucleotide variants in gene bodies plus 150 million indels.
October 8th, 2026
Illumina launches BioInsight platform with SpliceAI2 included
Product launch

Ahead of the American Society of Human Genetics 2026 meeting, Illumina announced the BioInsight platform as a single environment for its analysis applications, AI tools, and curated datasets. SpliceAI2 is among the listed tools, along with BaseSpace and Connected Multiomics.
October 2026
August 2025
Study documents errors in precomputed SpliceAI scores
Research

A medRxiv preprint found annotation errors in 8.34% of theoretical SNVs in SpliceAI's precomputed scores, potentially missing clinically relevant variants.
August 27th, 2025
2019
Original SpliceAI released
Product launch

Illumina released SpliceAI, a deep learning model for predicting splice effects from DNA sequence, with precomputed scores for all theoretical SNVs.
2019

Scenarios

SpliceAI2 becomes the clinical standard for splice variant interpretation

Likely Resolves by End of 2027

Discussed by: Illumina, ClinGen

ClinGen already incorporates the original SpliceAI into its splice variant interpretation guidelines. If it adopts SpliceAI2, clinical labs would update their pipelines, and more cryptic splice variants would get classified as pathogenic. Illumina is positioning SpliceAI2 as the successor, with precomputed scores for 4 billion SNVs and 150 million indels.

Independent validation finds SpliceAI2 limitations

Possible Resolves by End of 2027

Discussed by: Academic researchers, medRxiv preprint authors

A 2025 medRxiv paper documented annotation and liftover errors in the original SpliceAI's precomputed scores, affecting 8.34% of theoretical SNVs. Independent groups may find similar issues in SpliceAI2's precomputed predictions, which would slow clinical adoption.

Competing models close the gap

Possible Resolves by End of 2027

Discussed by: Google DeepMind, academic labs

Google DeepMind's AlphaGenome and other models compete in splice prediction. SpliceAI2 outperformed AlphaGenome in benchmarks run by University of Oxford collaborators, but DeepMind could release an updated model. If a rival reports better performance on independent benchmarks, SpliceAI2's advantage narrows.

Exome-negative patients get diagnoses from deep-intronic variants

Possible Resolves by Oct 1, 2027

Discussed by: Illumina, Genomics England collaborators

About half of the cryptic splice variants SpliceAI2 flagged sat more than 50 base pairs into introns, where exome sequencing typically misses them. If labs reanalyze existing genome data with SpliceAI2, some patients who tested negative could get a diagnosis.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

November 2020 - July 2021

AlphaFold (2020-2021)

DeepMind released AlphaFold, an AI model that predicted protein structures with near-experimental accuracy, solving a 50-year problem in biology. The 2021 version, AlphaFold2, was described as a breakthrough by the journal Nature.

Then

AlphaFold became widely adopted, with millions of protein structures predicted and shared openly through the AlphaFold Database.

Now

It transformed structural biology and demonstrated that deep learning could solve fundamental problems in life sciences.

Why this matters now

Like AlphaFold, SpliceAI2 applies deep learning to a fundamental biological problem. Both show how AI models trained on large datasets can become standard tools in genomics.

January 2019

Original SpliceAI (2019)

Illumina's AI lab released SpliceAI, a deep learning model that predicted splice effects from DNA sequence. It became the standard tool for splice variant interpretation, cited in over 3,400 publications.

Then

SpliceAI was incorporated into ClinGen guidelines and variant annotation tools like VEP and dbNSFP.

Now

It became the default choice for splice prediction in clinical genomics, though a 2025 medRxiv paper documented errors in its precomputed scores.

Why this matters now

SpliceAI2 is the direct successor, trained on a dataset 100 times larger and designed to address the original model's limitations.

2013-2018

100,000 Genomes Project (2013-2018)

Genomics England sequenced 100,000 genomes from NHS patients with rare diseases and cancer, aiming to create a national genomic medicine service. The project built sequencing infrastructure and identified diagnoses for thousands of patients.

Then

The project demonstrated that whole-genome sequencing could find diagnoses in rare disease, but also exposed the interpretation bottleneck.

Now

It showed that sequencing data alone is not enough - interpreting variants remains the hardest part.

Why this matters now

SpliceAI2 was validated on Genomics England data, and its purpose is to improve the interpretation bottleneck that the project exposed.