Predicts splice-site usage, splice junctions and full transcript isoforms from 196 kb of DNA, scoring how variants quantitatively alter splicing.
No providers recorded yet. Browse all providers
A variant that strengthens a cryptic acceptor ten nucleotides from a real one does not simply "create a splice site." It shifts some fraction of transcripts to a different exon boundary, and whether that matters depends on how much usage moves and which isoforms result. The original SpliceAI, released by Illumina in 2019, answered the first question well: will this position be spliced or not? It did not say how splice sites pair up into introns or which transcripts a gene ends up producing.
SpliceAI2 is the successor, described in a 2026 bioRxiv preprint by Kishore Jaganathan, Kyle Kai-How Farh and colleagues at Illumina, with collaborators at the University of Oxford, UCSF and the New York Genome Center. It predicts three linked quantities from DNA alone: how often each position is used as a donor or acceptor, how often each pair of candidate sites is joined as a junction, and which complete transcripts those sites and junctions assemble into. Candidate sites become nodes in a splice graph, junctions become edges, and each transcript is a scored path. Dynamic programming makes training against observed transcript usage tractable over the exponentially many paths.
A single trained model is applied without retraining to variant benchmarks and clinical cohorts. Its niche overlaps with OpenSpliceAI, a Johns Hopkins reimplementation of the original SpliceAI, and with the splicing heads of AlphaGenome, and neither of those predicts complete transcript isoforms.
SpliceAI2 reads 196,608 bp of sense-strand genomic sequence and predicts donor and acceptor usage across the central 65,536 positions. A dense projection to 256 channels feeds 20 MetaFormer blocks, each pairing a depth-wise dilated convolution with a feed-forward channel mixer at progressively coarser pooled resolution. The 1,024 highest-scoring positions are then passed to a query-key head with rotary position embeddings that scores every upstream-downstream pair as a candidate junction. Each model has about 13 million parameters, and the release averages predictions from two independently trained checkpoints. Training used splice site and junction counts from 314,745 bulk RNA-seq samples across ten mammalian species (recount3 and the Sequence Read Archive), plus transcript usage from 330 ENCODE long-read RNA-seq samples in human and mouse.
On variant benchmarks that exclude essential splice variants, SpliceAI2 reached an auPRC of 0.77 against 0.66 for SpliceAI in detecting cryptic splice activation among 4,545 GTEx variants. It reached an auROC of 0.76 against 0.73 for AlphaGenome in classifying fine-mapped GTEx sQTLs, and a Spearman ρ of 0.59 against 0.55 for SpliceAI on the OpenSplice MPRA. In 36,764 UK Biobank participants, SpliceAI2 loss scores for rare splice-region variants correlated with carriers' plasma protein levels (Pearson r = −0.50). All results come from a preprint that has not yet been peer reviewed.
SpliceAI2 targets non-coding variant interpretation in whole-genome sequencing, especially rare disease diagnosis. In 7,504 probands from the Genomics England 100,000 Genomes Project, SpliceAI2-prioritized variants accounted for an estimated 15% of the excess genetic burden in phenotype-matched genes, and about half lay more than 50 bp into introns, where exome sequencing typically does not reach. Blood RNA-seq from 5,435 participants confirmed predicted splicing changes, including a disrupted splice site in PEX1 and a cryptic site in PKD1. Because it works from DNA alone, it also covers genes not expressed in accessible tissues such as blood. Precomputed scores cover all 4 billion possible SNVs and 150 million population indels in human protein-coding genes.
The original SpliceAI is built into clinical guidelines for non-coding variant interpretation, so its successor arrives with an established user base and an upgrade path through compatible score columns. SpliceAI2 moves sequence-based splicing models from site classification toward full isoform prediction. Tissue-specific variant effects remain open: as with AbSplice2, reference splicing differences between tissues predicted them as well as tissue-conditioned predictions did. Its reach is limited by licensing: the code, weights and precomputed scores are released under Illumina's own terms of use, restricted to academic and non-commercial research use, with commercial use requiring a separate license from Illumina.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.