Bio Splice Variant Prediction

Version Compatibility

Reference examples tested with: SpliceAI 1.3+, Pangolin 1.0+, MMSplice 2.4+, pyensembl 2.3+, pysam 0.22+, pandas 2.2+, gffutils 0.13+, tensorflow 2.15+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Splice Variant Prediction

Predict whether a DNA variant alters mRNA splicing. Distinct from "variant pathogenicity" generally: a variant can be a strong splice disruptor without being pathogenic for the gene's standard mechanism, or pathogenic for reasons orthogonal to splicing. Splice prediction asks specifically: does this variant change splice-site usage?

Predictor Taxonomy

Tool Selection Matrix

Methodology evolves; verify benchmarks (Smith & Kitzman 2023 Genome Biol 24:294; You et al 2024 Nat Commun) and ClinGen SVI splicing recommendations before reporting clinical interpretations. Concordance across SpliceAI + Pangolin + MMSplice is gold-standard evidence; discordance flags need RNA validation.

Decision Tree by Use Case

ClinGen SVI 2023 Framework

The ClinGen Sequence Variant Interpretation (SVI) splicing subgroup (Walker 2023 Am J Hum Genet) extended the ACMG/AMP 2015 framework with explicit splice-prediction rules.

Operational rules: Computational evidence (PP3/BP4) is supporting, not standalone. ClinGen SVI 2023 recommends applying predictive splice PP3/BP4 at supporting weight only; higher SpliceAI cutoffs (0.5, 0.8) increase precision but are the tool's own tiers (Jaganathan 2019), NOT ClinGen-endorsed evidence-strength upgrades — reaching moderate/strong requires functional/RNA evidence (PS3/BS3), not a higher SpliceAI score alone. Splicing variants benefit from concordance across SpliceAI + Pangolin + MMSplice. RNA validation supersedes prediction. Always log SpliceAI version, distance window, and reference transcript. SpliceAI alone is not sufficient for PVS1; canonical site disruption requires gene-level LoF context.

SpliceAI Workflow

Goal: Annotate VCF variants with per-variant delta scores for splice-site change.

Approach: Run spliceai CLI with reference genome and annotation; parse INFO field for delta scores. SpliceAI is human-only (-A grch37 or -A grch38); the model was trained on GENCODE human and does not directly transfer to mouse, fly, or other species. For mouse, retrained variants exist (e.g. mouseSpliceAI); for other species, use Pangolin (4 species: human, mouse, rat, rhesus macaque) or accept that prediction will be unreliable.

spliceai \
    -I input.vcf \
    -O output.vcf \
    -R GRCh38.primary_assembly.genome.fa \
    -A grch38 \
    -D 50 \
    -M 0

-D 50 = distance window in nt around variant (default 50). For deep-intronic variants suspected of creating pseudoexons, raise to 500-2000:

spliceai -I input.vcf -O output_extended.vcf -R genome.fa -A grch38 -D 500 -M 1

-M 0 (default) returns raw scores; -M 1 masks splice gains at annotated sites and losses at unannotated sites (cleaner for clinical use). Output INFO format: SpliceAI=ALLELE|SYMBOL|DS_AG|DS_AL|DS_DG|DS_DL|DP_AG|DP_AL|DP_DG|DP_DL. Delta score = max(DS_AG, DS_AL, DS_DG, DS_DL).

import pandas as pd
import re

def parse_spliceai_vcf(vcf_path):
    rows = []
    with open(vcf_path) as f:
        for line in f:
            if line.startswith('#'):
                continue
            fields = line.strip().split('\t')
            info = fields[7]
            m = re.search(r'SpliceAI=([^;]+)', info)
            if not m:
                continue
            for ann in m.group(1).split(','):
                parts = ann.split('|')
                allele, symbol = parts[0], parts[1]
                ds = [float(p) if p != '.' else 0 for p in parts[2:6]]
                dp = parts[6:10]
                rows.append({
                    'chrom': fields[0], 'pos': int(fields[1]),
                    'ref': fields[3], 'alt': allele,
                    'gene': symbol,
                    'DS_AG': ds[0], 'DS_AL': ds[1],
                    'DS_DG': ds[2], 'DS_DL': ds[3],
                    'delta_max': max(ds),
                })
    return pd.DataFrame(rows)

df = parse_spliceai_vcf('output.vcf')
df['acmg_evidence'] = pd.cut(
    df['delta_max'],
    bins=[-0.01, 0.10, 0.20, 0.50, 0.80, 1.01],
    # ClinGen SVI applies splice PP3/BP4 at supporting weight; 0.5/0.8 are SpliceAI
    # precision tiers (Jaganathan 2019), NOT ACMG evidence-strength upgrades
    labels=['BP4', 'inconclusive', 'PP3_supporting', 'PP3_supporting_prec0.5', 'PP3_supporting_prec0.8']
)

DS labels: AG = acceptor gain, AL = acceptor loss, DG = donor gain, DL = donor loss.

Pangolin for Tissue-Specific Prediction

Goal: Get tissue-specific splice impact predictions when disease tissue is known.

Approach: Run Pangolin CLI with VCF + reference + gffutils annotation database.

python -c "import gffutils; gffutils.create_db('gencode.v45.annotation.gff3', 'gencode.db', force=True)"

pangolin \
    input.vcf \
    GRCh38.primary_assembly.genome.fa \
    gencode.db \
    pangolin_output \
    -d 500 \
    -m True \
    -s 0.2

-m True masks splice gains at annotated sites and losses at unannotated sites (recommended for clinical use). -s 0.2 outputs all sites with predicted change >= cutoff.

Pangolin output is a VCF with per-tissue predictions across the 4 tissues used at training: brain, heart, liver, testis (Zeng & Li 2022 Genome Biol). The model outputs per-species per-tissue predictions but extrapolates poorly to tissues outside this set. Use the tissue closest to disease-relevant context. For tissues not in the 4-tissue training set, fall back to SpliceAI — Pangolin extrapolates poorly to unseen tissues.

SpliceVault for Empirical Mis-Splicing Outcomes

Goal: Predict the type of mis-splicing (exon skipping vs cryptic site activation) given a canonical-disrupting variant.

Approach: Query SpliceVault's database of empirical mis-splicing events from public RNA-seq.

import requests

# Web API: https://kidsneuro.shinyapps.io/splicevault/
# Or use the R/Python package at github.com/kidsneuro-lab/SpliceVault

# Example: NM_000546.6:c.673-2A>G (TP53)
# Returns top-N most likely mis-splicing events: exon skipping, cryptic 3'ss usage, etc.

SpliceVault (Dawes 2023 Nat Genet) showed that the Top-4 events at any splice site predict variant-associated mis-splicing with ~92% sensitivity overall (96% of exon-skipping and 86% of cryptic-activation events) — a striking regularity that makes consequence prediction tractable. Use SpliceVault when the question is not "will splicing change?" but "what specific aberrant splicing will occur?".

MMSplice for Calibrated ΔPSI

Goal: Predict quantitative ΔPSI (not just probability of disruption) for cassette exons.

Approach: Score variant impact on each splicing region (5'ss, 3'ss, exon, intron-3'/5') and combine.

from mmsplice.vcf_dataloader import SplicingVCFDataloader
from mmsplice import MMSplice, predict_save

dl = SplicingVCFDataloader(
    gtf='gencode.v45.basic.gtf',
    fasta_file='GRCh38.fa',
    vcf_file='input.vcf'
)

model = MMSplice()
predict_save(model, dl, 'mmsplice_predictions.csv', pathogenicity=True)

MMSplice (Cheng 2019 Genome Biol) reports Δlogit_psi per variant. Useful when calibrated effect sizes matter (research) more than probability of disruption (clinical screening). Companion MTSplice (Cheng 2021 Genome Biol) adds tissue-specific Δψ predictions.

HGVS Splicing Nomenclature

Following den Dunnen 2016 Hum Mutat:

Validation tools: VariantValidator (Freeman 2018 Hum Mutat), Mutalyzer 2 (Lefter et al 2021 Bioinformatics 37:2811-2817).

Extended-Window Scoring for Deep-Intronic Variants

SpliceAI's default precomputed scores use a 50-nt window, missing variants that create pseudoexons in deep intronic regions. For unsolved Mendelian cases:

# Recompute with extended window
spliceai -I input.vcf -O output_2kb.vcf -R genome.fa -A grch38 -D 2000

# Or use CI-SpliceAI (Strauch 2022 PLoS One), SpliceAI retrained on curated GENCODE splice sites

Pseudoexon creation in deep introns explains a substantial fraction of unsolved Mendelian disease alleles in current cohorts (estimates 5-15% across studies; specific quantitative range will vary by cohort and panel — verify against current literature). Disease examples: CFTR 3849+10kbC>T, USH2A c.7595-2144A>G, CEP290 c.2991+1655A>G (LCA10).

Concordance Across Predictors

import pandas as pd

merged = (spliceai_df
    .merge(pangolin_df, on=['chrom', 'pos', 'alt'], suffixes=('_sai', '_pang'))
    .merge(mmsplice_df, on=['chrom', 'pos', 'alt'])
)

merged['concordance'] = (
    (merged['delta_max_sai'] >= 0.2).astype(int) +
    (merged['pangolin_score'].abs() >= 0.2).astype(int) +
    (merged['delta_logit_psi'].abs() >= 1.0).astype(int)
)

merged['interpretation'] = merged['concordance'].map({
    0: 'concordant_benign',
    1: 'discordant_low_evidence',
    2: 'concordant_evidence',
    3: 'high_concordance_pathogenic'
})

Discordance is the most informative pattern — variants where one model sees impact and others don't are high priority for RNA validation.

Branchpoint Variant Detection

All current tools are weak at branchpoint variants because the BPS motif (yUnAy) has low information content. Specific branchpoint tools:

Branchpoint variants are under-recognized in clinical pipelines; SpliceAI captures only some because branchpoint motifs have low information content. Recommendation: when SpliceAI delta is borderline (0.1-0.3) for a variant in the BPS region (-18 to -40 from 3'ss), run BPHunter as supplement.