nicolynnvila/splicing_epistasis_okgp_test

Splicing epistasis — 1000 Genomes matched-arms test set

End-to-end reproducibility test set for the splicing-epistasis paper (Lynn et al., Genome Medicine, in preparation). 28,642 cis-double SNV pairs from a 138-gene cancer panel, matched in two arms (observed in 1000 Genomes vs. matched unobserved control).

Test purpose

This dataset exists to validate the engine-inference layer of the paper's pipeline. The 4-engine splice-prediction ensemble (Pangolin, SpliceAI-PyTorch, Spliceformer, OpenSpliceAI) was originally scored on a private compute cluster; this test set lets anyone re-score it on portable infrastructure (e.g. Modal, RunPod, or local GPU) and compare against the ground-truth predictions.

Composition

• n = 28,642 pairs
• arms: 14,321 observed (both variants co-occur in ≥ 1 1KG individual) + 14,321 matched unobserved (both variants exist individually in 1KG but never together; 1:1 matched on gene, log-rarity bin, and inter-variant distance ± 2 nt)
• panel: 132 genes (of a 138-gene cancer panel; 6 genes had no scoreable pair)
• distance: 1–100 nt (within-gene, intra-intron-window)

Files

• inputs.parquet — the 28,642 pairs with constituent variants, distance, gene, intron positions, and ClinVar metadata. This is the input you score.

Columns (inputs)

Column	Type	Description
epistasis_id	str	GENE:CHR:POS:REF:ALT|GENE:CHR:POS:REF:ALT canonical pair ID
arm	str	observed (both variants co-occur in 1KG) or unobserved (matched control)
gene	str	HGNC gene symbol
chrom	str	chromosome
mut1, mut2	str	constituent variant IDs (GENE:CHR:POS:REF:ALT)
distance	int	inter-variant distance (nt)
n_mut1, n_mut2	int	1KG carrier count per constituent variant
n_both	int	1KG carriers with BOTH variants (always 0 for unobserved arm)
intron_donor_pos, intron_acceptor_pos	int	nearest annotated splice-site coordinates
dataset	str	provenance tag
clinsig_1, clinsig_2	str	ClinVar pathogenicity per constituent (if annotated)
disease_1, disease_2	str	ClinVar disease per constituent (if annotated)
shared_disease	bool	whether the two ClinVar disease annotations match
both_clinvar_annotated	bool	whether both constituents are in ClinVar
gene_type	str	tumor-suppressor / oncogene / other (from OncoKB)

Quick start

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/splicing_epistasis_okgp_test")
df = ds["train"].to_pandas()
print(df.head())
print(f"{len(df):,} pairs, {df.arm.value_counts().to_dict()}")

Scoring contract

A scorer (any splice-prediction engine) should consume the input pairs and produce, per pair:

• psi_ref — reference PSI at the focal junction
• psi_event — predicted PSI under the cis-double (mut1 + mut2 in cis)
• psi_mut1, psi_mut2 — predicted PSI under each single variant alone
• psi_expected = psi_mut1 + psi_mut2 - psi_ref (additive null)
• psi_residual = psi_event - psi_expected (the epistatic signal)

The paper uses |psi_residual| ≥ 0.10 as the call threshold under a 1-of-4 ensemble rule (any one of 4 engines passing).

Paper reproduction target

For a successful end-to-end test, scoring this 28k set with OpenSpliceAI should reproduce the psi_residual_openspliceai column from the paper's bundled master parquet within numerical tolerance.

Citation

@article{lynn2026splicing,
  title  = {Splicing-context epistasis identifies cis-double variants with
            joint splice-disrupting effects across health and cancer},
  author = {Lynn, Nicolas A. and ... and Pe'er, Itsik},
  journal = {Genome Medicine},
  year   = {2026}
}

Source code + full pipeline

Project repository (in progress): https://github.com/nicolaslynn/splicing_epistasis