SpliceAI

SpliceAI processes 10,000 nucleotides of genomic context flanking each position, dramatically expanding the receptive field beyond traditional splice site predictors. This long-range window captures distal branch points, polypyrimidine tracts, and splicing regulatory elements that influence exon inclusion.

• Uses dilated convolutions with exponentially increasing dilation rates
• Receptive field spans 10 kb, covering most human introns
• Detects cryptic splice sites activated by deep intronic variants
• Outperforms window-based methods limited to 200-400 nucleotides

The delta score quantifies the functional impact of a genetic variant on splicing by computing the difference between predicted splice site probabilities for the mutant and reference sequences. A higher delta score indicates a stronger likelihood that the variant disrupts normal splicing.

• Ranges from 0 to 1, with 0.5 as a commonly used threshold for high-impact predictions
• Identifies cryptic splice site activation in deep intronic regions
• Clinically validated for neurodevelopmental disorders and rare disease diagnostics
• Enables prioritization of non-coding variants of unknown significance

SpliceAI simultaneously predicts splice donor (5' splice site) and splice acceptor (3' splice site) probabilities at every position in the input sequence. The model outputs position-specific scores for each motif, enabling comprehensive annotation of both annotated and novel splice junctions.

• Predicts canonical GT-AG and non-canonical splice site pairs
• Identifies alternative splicing events including exon skipping and intron retention
• Detects tissue-specific splice sites when trained with appropriate labels
• Generates per-position probability tracks for genome-wide annotation

The model employs a deep residual architecture with 32 convolutional layers and skip connections that enable gradient flow through the network during training. Each residual block contains dilated convolutions that exponentially increase the receptive field without adding parameters.

• 32 dilated convolutional layers with batch normalization and ReLU activations
• Skip connections every 2 layers to prevent vanishing gradients
• Exponential dilation pattern: 1, 2, 4, 8, 16, 32, 64, 128, 256, 512
• Trained on GENCODE-annotated splice sites from the human reference genome

SpliceAI accepts raw pre-mRNA nucleotide sequences as input, encoding each position as a one-hot vector over the four canonical bases (A, C, G, T) plus an ambiguous base (N). No hand-crafted features or conservation scores are required, allowing the model to learn splicing motifs de novo.

• Input shape: 4-channel one-hot encoding across 10,000 positions
• No reliance on phyloP or GERP conservation scores
• Learns splicing regulatory motifs directly from sequence
• Applicable to any species with annotated transcriptomes for retraining

SpliceAI has been integrated into clinical genomics pipelines to prioritize non-coding variants and synonymous coding variants that may disrupt splicing. The model's delta scores are used alongside population frequency and inheritance data to identify pathogenic variants in rare disease cohorts.

• Validated in autism spectrum disorder and intellectual disability studies
• Identifies deep intronic variants missed by exome sequencing analysis
• Integrated into gnomAD and ClinVar variant annotation resources
• Complements in silico mutagenesis approaches for saturation mutagenesis screening

SpliceAI is integrated into clinical pipelines to resolve variants of uncertain significance (VUS) by predicting whether single nucleotide variants create cryptic splice sites or disrupt canonical donor/acceptor motifs.

• Predicts delta scores for acceptor gain, acceptor loss, donor gain, and donor loss at any position
• A delta score > 0.5 indicates a high-confidence splicing alteration
• Used alongside ACMG/AMP guidelines to upgrade variant pathogenicity classifications
• Processes a 10,000-nucleotide window to capture deep intronic splicing regulatory elements

Whole-genome and whole-exome sequencing often identifies non-coding variants whose functional impact is unclear. SpliceAI systematically scans these variants to uncover cryptic splicing defects underlying Mendelian disorders.

• Identifies deep intronic variants that create pseudoexons through cryptic splice site activation
• Applied in large-scale rare disease cohorts such as the 100,000 Genomes Project
• Enables reanalysis of previously negative exomes by flagging missed splicing-altering variants
• Complements RNA-seq validation by prioritizing variants for functional follow-up

Researchers use SpliceAI to probe the cis-regulatory logic of the spliceosome by systematically perturbing sequences in silico and observing predicted splicing changes.

• Enables high-throughput in silico mutagenesis across entire gene loci
• Reveals positional biases in exonic splicing enhancers (ESEs) and silencers (ESSs)
• Maps the sequence determinants of alternative splicing across tissue types
• Validates experimental findings from Massively Parallel Reporter Assays (MPRAs)

In precision medicine, SpliceAI aids the discovery of druggable splicing targets by predicting which variants create splice isoforms amenable to antisense oligonucleotide (ASO) or small molecule correction.

• Identifies variants that induce exon skipping or cryptic exon inclusion correctable by ASOs
• Supports the development of splice-switching therapies for neuromuscular disorders
• Predicts off-target splicing effects of CRISPR base editors and prime editors
• Used in preclinical screening for nonsense-mediated decay (NMD) evasion mechanisms

By applying integrated gradients and in silico mutagenesis to SpliceAI, researchers extract the sequence motifs driving splicing predictions, revealing the model's learned biological logic.

• Recovers known 5' donor (GT) and 3' acceptor (AG) dinucleotide motifs without explicit training
• Identifies branch point sequences and polypyrimidine tract signals learned from context
• Generates saliency maps that highlight nucleotides critical for exon definition
• Validates that the model has internalized established spliceosome biochemistry