SpliceAI

Predict the effect of genetic variants on RNA splicing.

• Paper: Cell 2019 — Jaganathan et al., "Predicting Splicing from Primary Sequence with Deep Learning"
• Upstream: github.com/Illumina/SpliceAI
• License: Non-commercial only. Source under PolyForm Strict 1.0.0; bundled weights under CC-BY-NC-4.0. Commercial use requires a license from Illumina. RNAZoo distributes the upstream pip package in the image; downstream users are bound by the same terms.
• Device: CPU or GPU. Two image variants:
  • rnazoo-spliceai:latest — CUDA-enabled (default, used with -profile gpu)
  • rnazoo-spliceai-cpu:latest — CPU-only (smaller, used with -profile cpu)

License heads-up: SpliceAI is the only model in the zoo with non-commercial restrictions on both source and weights. If you are using RNAZoo in a commercial context (drug-development pipelines, paid clinical interpretation, etc.) you must either obtain a commercial license from Illumina or skip this model (don't pass --spliceai_vcf).

What it does

SpliceAI is a 32-layer dilated convolutional neural network that takes a candidate variant (SNV or short indel) plus its surrounding ±5 kb of reference sequence and predicts the variant's effect on splicing. For each variant it emits four delta scores in [0, 1]:

Score	Meaning
DS_AG	Acceptor-gain delta
DS_AL	Acceptor-loss delta
DS_DG	Donor-gain delta
DS_DL	Donor-loss delta

…plus the relative position (DP_*) of each putative gain/loss event. The recommended delta-score threshold for a splice-altering variant is 0.5; the paper uses 0.2 (high recall) and 0.8 (high precision) as alternative cutoffs.

The pipeline currently exposes only the inference path. Upstream's training scripts and supplementary analyses are not wired in.

Input format

Three inputs are required:

1. VCF — variants of interest, with CHROM, POS, REF, ALT columns. Multi-allele records are supported; complex SVs and indels longer than 2 × distance are silently skipped.
2. Reference FASTA — must contain every chromosome referenced in the VCF. A .fai sibling is created on the fly by pyfaidx if missing.
3. Annotation — one of:
   • grch37 (bundled GENCODE V24lift37 canonical TSV)
   • grch38 (bundled GENCODE V24 canonical TSV)
   • a path to a custom GENCODE-style TSV with columns #NAME CHROM STRAND TX_START TX_END EXON_START EXON_END (trailing-comma-separated exon coordinate lists)

Variants outside any gene defined in the annotation are silently skipped — SpliceAI scores nothing for intergenic positions.

Example VCF:

##fileformat=VCFv4.2
##contig=<ID=chr19,length=58617616>
#CHROM  POS       ID  REF  ALT  QUAL  FILTER  INFO
chr19   38958362  .   C    T    .     .       .

Output format

An annotated VCF where each variant gains a SpliceAI=... field in the INFO column:

chr19  38958362  .  C  T  .  .  SpliceAI=T|RYR1|0.00|0.00|0.91|0.08|-28|-46|-2|-31

Format: ALLELE|SYMBOL|DS_AG|DS_AL|DS_DG|DS_DL|DP_AG|DP_AL|DP_DG|DP_DL.

Run with Docker

See the Direct Docker guide for the shared docker run recipe (UID, HOME, USER env vars, and GPU flag). Below are the model-specific parts.

# CPU
docker run --rm \
  -v /path/to/inputs:/work \
  ghcr.io/ericmalekos/rnazoo-spliceai-cpu:latest \
  spliceai_predict.py \
    -i /work/variants.vcf \
    -o /work/variants.annot.vcf \
    -r /work/genome.fa \
    -a grch38

# GPU
docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all \
  -v /path/to/inputs:/work \
  ghcr.io/ericmalekos/rnazoo-spliceai:latest \
  spliceai_predict.py \
    -i /work/variants.vcf \
    -o /work/variants.annot.vcf \
    -r /work/genome.fa \
    -a grch38

For a custom annotation TSV, replace -a grch38 with -a /work/my_annot.tsv and bind-mount the file into /work.

Run with Nextflow

# CPU, builtin GRCh38 annotation
nextflow run main.nf -profile docker,cpu \
  --spliceai_vcf /path/to/variants.vcf \
  --spliceai_reference_fasta /path/to/genome.fa \
  --spliceai_annotation grch38

# GPU, custom annotation TSV
nextflow run main.nf -profile docker,gpu \
  --spliceai_vcf /path/to/variants.vcf \
  --spliceai_reference_fasta /path/to/genome.fa \
  --spliceai_annotation /path/to/my_annot.tsv

Only models with input provided will run — no ignore flags needed.

Results appear in results/spliceai/spliceai_out/<basename>.spliceai.vcf.

Parameters

Parameter	Default	Description
--spliceai_vcf	(required)	Input VCF with variants of interest
--spliceai_reference_fasta	(required)	Reference FASTA matching the VCF coordinates; .fai autogenerated if absent
--spliceai_annotation	grch38	grch37, grch38, or path to a custom GENCODE-style TSV
--spliceai_distance	50	Maximum distance (bp) between the variant and any predicted gain/loss site
--spliceai_mask	0	0 = raw scores; 1 = mask annotated-acceptor/donor gain and unannotated loss (recommended for variant interpretation)

Reading the output

import pysam

vcf = pysam.VariantFile("variants.annot.vcf")
for rec in vcf:
    info = rec.info.get("SpliceAI")
    if not info:
        continue
    for entry in info:
        allele, gene, ds_ag, ds_al, ds_dg, ds_dl, *_ = entry.split("|")
        deltas = [float(x) for x in (ds_ag, ds_al, ds_dg, ds_dl)]
        max_delta = max(deltas)
        if max_delta >= 0.5:
            print(f"{rec.chrom}:{rec.pos} {rec.ref}>{allele} ({gene}) — likely splice-altering ({max_delta:.2f})")

Test fixture (synthetic mini-genome)

The bundled smoke fixture under tests/data/spliceai/ is a synthetic 12-kb chromosome (TEST_CHR) with random ACGT, a custom annotation defining a fake TEST_GENE with 3 exons, and a single SNV at the exon-2 midpoint. Delta scores from this fixture are all 0.00 — the random sequence has no real splice motifs — but the test verifies that the pipeline plumbing works end-to-end. For real predictions, use a real reference genome (GRCh38 from UCSC or Ensembl) and the bundled grch38 annotation.

Limitations

• Non-commercial only. See the License section above.
• Variants must be inside annotated genes. Intergenic variants are silently skipped; if your VCF appears empty in the output, check that the variants overlap genes in your annotation.
• Variants close to chromosome ends are skipped (within 5 kb of either end — the model's per-variant context window can't be filled).
• Indels longer than 2 × distance are skipped. Bump --spliceai_distance if you need to score larger indels (default 50 → indels up to 100 bp).
• Reference allele is checked. If REF doesn't match the reference FASTA at POS, SpliceAI emits a warning and skips that record.