Skip to content

SpliceAI

Summary

SpliceAI predicts, for every position in a pre-mRNA transcript, whether it acts as a splice donor, a splice acceptor, or neither — using 10,000 nucleotides of flanking genomic sequence and nothing else.[1] Variants are scored by the delta (Δ) score: the largest change in predicted splice probability within 50 nt of the variant, comparing reference and alternate sequence. Because the network was trained only on reference transcripts and annotations and never saw variant data, its variant predictions are a genuine generalisation test rather than a fitted task — a design decision that makes SpliceAI's clinical adoption defensible in a way most variant predictors are not.[1]

Architecture

  • SpliceAI-10k: 32 dilated convolutional layers in a deep residual network, taking 5,000 nt of flanking context on each side of the position being scored.[1]
  • Smaller variants exist for ablation and speed: SpliceAI-80nt, -400nt and -2k use 40, 200 and 1,000 nt of flanking sequence per side.[1]
  • Trained on GENCODE-annotated pre-mRNA transcript sequences from a subset of human chromosomes; tested on the remaining chromosomes with paralogs excluded.[1]

What the model learned

This is the part of the paper most often overlooked, and it is the reason SpliceAI outperformed motif-based predictors (GeneSplicer, MaxEntScan, NNSplice) so decisively.[1]

  • Top-k accuracy of 95% on held-out transcripts; genes over 100 kb such as CFTR are often reconstructed to nucleotide precision.[1]
  • 84% top-k accuracy on lincRNAs, confirming the network is not simply exploiting protein-coding exonic sequence bias — it approximates spliceosome behaviour on sequences free of coding selective pressure.[1]
  • Prediction scores track exon inclusion rate measured in GTEx (Pearson 0.78): constitutively included or excluded exons score near 1 or 0, alternatively spliced exons score in between.[1]
  • Long-range determinants dominate. Increasing training context from 80 nt to 10 kb markedly improves accuracy. The 80 nt model assigns lower scores to junctions of typical exon/intron length, whereas the 10 kb model prefers them — because it can account for exon and intron length as specificity signals. The authors' interpretation: the apparent degeneracy of most splice motifs is explained by compensating long-range context, which is why motif-only scoring has poor specificity genome-wide.[1]
  • The network implicitly learned nucleosome positioning. Walking an optimal acceptor/donor pair 150 nt apart across the genome, predicted exon-inclusion favourability correlated with nucleosome occupancy (Spearman 0.36), persisting after GC control. Private exon-creation events in GTEx were significantly associated with pre-existing nucleosome positioning in cell lines that lack the private mutation (p = 0.006), supporting a causal role for nucleosome positioning in exon definition.[1]

Validation: three orthogonal lines of evidence

RNA-seq. In 149 GTEx individuals with matched WGS and RNA-seq, predicted splice-altering private variants are strongly enriched at private novel junctions and skipped-exon boundaries. Confidently predicted cryptic splice variants (Δ ≥ 0.5) validate at three-quarters the rate of essential GT/AG disruptions, and both validation rate and effect size track the Δ score.[1] Sensitivity at Δ ≥ 0.5 is 71% near exons but only 41% for deep intronic variants — deep intronic prediction is the model's weak axis.[1]

Natural selection. Scoring 60,706 ExAC exomes and restricting to synonymous and intronic variants (to avoid coding confounders), predicted cryptic splice variants at Δ ≥ 0.8 are depleted 78% at common allele frequencies (odds ratio 4.58, p < 10⁻¹²⁷) — comparable to the 82% depletion of frameshift, stop-gain and essential-splice variants.[1] Extending to deep intronic regions using 15,496 gnomAD genomes gives 56% depletion, consistent with the lower deep-intronic accuracy.[1]

Case–control de novo enrichment. Across 4,293 DDD intellectual disability patients, 3,953 autism cases and 2,073 sibling controls, predicted splice-disrupting de novo mutations are enriched 1.51-fold in intellectual disability (p = 0.000416) and 1.30-fold in autism (p = 0.0203) at Δ ≥ 0.1, implying that cryptic splice mutations account for ~9% of pathogenic de novo mutations in intellectual disability and ~11% in autism.[1] Including them in gene discovery yielded 5 additional candidate ID genes and 2 additional ASD genes at FDR < 0.01.[1]

Direct experiment. Deep RNA-seq (~350 million × 150 bp reads, ~10× GTEx coverage) of lymphoblastoid lines from 36 Simons Simplex Collection individuals: after excluding 8 for insufficient coverage, 21 of 28 (75%) validated — 9 novel junction creations, 8 exon skips, 4 intron retentions.[1]

Population burden and penetrance

  • The average human carries ~5 rare functional cryptic splice mutations versus ~11 rare protein-truncating variants; cryptic splice variants outnumber essential GT/AG disruptions roughly 2:1.[1]
  • Cryptic splice mutations have roughly 50% of the clinical penetrance of classic protein-truncating variants, because many only partially shift splicing — the case-control enrichment is 38% of the effect size seen for de novo PTVs.[1] This is the quantitative basis for treating a Δ-score-called pLoF as a weaker allele than a nonsense variant, even when effect sizes look comparable in a well-powered cohort.
  • Cryptic splice variants with weak-to-intermediate scores (Δ 0.35–0.8) frequently produce tissue-specific splicing: 35% show significant across-tissue differences in aberrant transcript fraction, whereas Δ > 0.8 variants are significantly less likely to be tissue-specific (p = 0.015).[1]

Disputed / unresolved: Koyama et al. (2026) report that cryptic splice variants called at Δ > 0.8 in a 1,158,017-individual lipid ExWAS have phenotypic effect sizes statistically indistinguishable from canonical pLoF variants (median β = 1.092, P = 0.71).[2] The SpliceAI primary paper estimates ~50% clinical penetrance relative to protein-truncating variants — but at the far more permissive Δ ≥ 0.1 threshold, and on de novo neurodevelopmental phenotypes rather than a quantitative blood lipid trait.[1] The two are plausibly reconciled by threshold and trait, not necessarily in conflict; neither study tested the other's design.

Usage in large-scale exome analysis

In the Koyama et al. (2026) blood-lipids ExWAS, variants with Δ > 0.8 were reclassified as predicted loss of function. The study identified 15 exome-wide significant cryptic splice variants, including a splice donor 5th-base variant in APOA5 (rs200831171) associated with elevated triglycerides.[2] See Cryptic Splice Variants.

Precomputed Δ scores for all possible single-nucleotide substitutions genome-wide are released as a community resource, which is why SpliceAI annotation is essentially free to add to any exome pipeline.[1]

Position in the splicing predictor landscape

SpliceAI is the short-context, base-resolution archetype of the sequence-to-function family. Its restrictions are structural: a ≤10 kb window, and modelling splicing at one level only — the probability that a nucleotide is a donor or acceptor — omitting competitive splice-site usage and splice junction identity.[3]

Newer models beat it on benchmark auPRC. On the MFASS minigene MPRA, SpliceAI and DeltaSplice both scored 0.49 versus 0.51 for AlphaGenome and 0.54 for Pangolin.[3] Pangolin reuses SpliceAI's ±5 kb receptive field almost unchanged and attributes its gain to training targets — quantitative, multi-tissue, multi-species splice-site usage rather than binary human annotations.[4] But three things keep SpliceAI in production use:

  1. A single, interpretable, thresholded score with clinically established cut-offs (0.2 / 0.5 / 0.8) and published validation rate and effect size at each.[1]
  2. Precomputed genome-wide tables, so annotation costs nothing at scale.[1]
  3. Orthogonal validation by RNA-seq, population selection and case-control enrichment — a standard of evidence most successor models have not reproduced.[1]
  • Cryptic Splice Variants — the variant class SpliceAI defined operationally.
  • Sequence-to-Function Genomic Models — the model family and the context/resolution trade-off SpliceAI sits on.
  • AlphaGenome — 1 Mb context, and splice junction prediction in addition to site probability.
  • Pangolin — the direct successor: same architecture family, quantitative tissue-specific training targets, better on every splicing benchmark tested.
  • Borzoi — an RNA-seq-coverage model that is complementary to splice-site models: better near junctions, worse on distant de novo splice gains.
  • Missense Score — deleteriousness metric used alongside splice-impact annotation in the lipid ExWAS.

Citations