mrunyan1/splaire-haec-sqtl-credible-sets

SPLAIRE scores for HAEC185 sQTL credible sets

SPLAIRE-ref and SPLAIRE-var variant effect predictions for all variants in the HAEC185 sQTL credible sets.

Files

• HAEC185_sQTL_credible_sets_splaire_scores.tsv.gz — per-credible-set-row summary (146,362 rows × 80 columns, 29 MB)
• h5/cs_all.splaire.ref.h5 — full ref/alt per-position scores from SPLAIRE-ref (86,549 unique variants, 26 GB)
• h5/cs_all.splaire.var.h5 — full ref/alt per-position scores from SPLAIRE-var (same variants, 26 GB)

The H5s hold the raw per-position score arrays the TSV is derived from.

Source data

146,362 variant-phenotype rows in the HAEC185 sQTL credible sets, deduplicated to 86,549 unique (variant, strand) pairs. Each pair scored with both SPLAIRE models. Input window 20,001 bp centered on each variant, 10,001 output positions covering ±5 kb.

Indel handling

Two-stage alignment so alt − ref comparisons stay position-aligned at the same genomic coordinate.

• Input window — the alt sequence is built by substituting at center, then padding (deletions) or trimming (insertions) the downstream end of the 20,001 bp window to keep total length fixed.
• Output realignment — the alt prediction track is realigned to ref coordinates. Extra positions from insertions are collapsed at center via element-wise max. Deleted positions are filled with zeros. Same convention as SpliceAI and Pangolin.

TSV columns

Shape: (146362, 80). First 8 columns from the original credible sets file:

gene, phenotype, ensembl_id, variant_id, credible_set_number, posterior_inclusion_probability, chr, strand

The remaining 72 are score columns templated by model (splaire, splaireVar) and head (don, acc, ssu) — 12 columns per (model, head):

column	meaning
{model}_{head}_max_inc	largest positive delta (alt − ref) in the 10 kb window
{model}_{head}_max_inc_off	offset from the variant (bp)
{model}_{head}_max_inc_pos	genomic coordinate of the max increase
{model}_{head}_max_dec	largest negative delta (alt − ref)
{model}_{head}_max_dec_off	offset from the variant (bp)
{model}_{head}_max_dec_pos	genomic coordinate of the max decrease
{model}_{head}_{coord}_ref	ref score at leafcutter junction boundary coord (istart = intron start, iend = intron end)
{model}_{head}_{coord}_alt	alt score at the same boundary
{model}_{head}_{coord}_delta	alt − ref at the same boundary

Intron-coordinate columns are NaN when the boundary falls outside the ±5 kb scored window (~87% of rows).

H5 contents

Each H5 contains five datasets (variant at center index 5000 of the 10,001-position window):

dataset	shape	dtype	meaning
cls_ref	(86549, 10001, 3)	float32	classifier softmax (neither, acceptor, donor), ref allele
cls_alt	(86549, 10001, 3)	float32	same, alt allele
reg_ref	(86549, 10001)	float32	regression SSU (sigmoid, 0-1), ref allele
reg_alt	(86549, 10001)	float32	same, alt allele
var_key	(86549,)	vlen UTF-8 string	variant identifier (chr:pos:ref:alt), sorted by (chrom, var_pos)

Use var_key to align H5 rows back to variant_id in the TSV.

Models

• splaire — SPLAIRE-ref, trained on reference genome sequences
• splaireVar — SPLAIRE-var, trained on sequences containing personal genetic variants

Each is a 5-fold ensemble, predictions are averaged across folds.

Output heads

• don — donor splice site probability (classifier softmax index 2)
• acc — acceptor splice site probability (classifier softmax index 1)
• ssu — splice site usage (regression sigmoid, 0-1)

Usage

import pandas as pd
import h5py
from huggingface_hub import hf_hub_download

tsv = hf_hub_download(
    repo_id="mrunyan1/splaire-haec-sqtl-credible-sets",
    filename="HAEC185_sQTL_credible_sets_splaire_scores.tsv.gz",
    repo_type="dataset",
)
df = pd.read_csv(tsv, sep="\t")
print(df.shape)  # (146362, 80)

# load full per-position scores for one model
h5_path = hf_hub_download(
    repo_id="mrunyan1/splaire-haec-sqtl-credible-sets",
    filename="h5/cs_all.splaire.ref.h5",
    repo_type="dataset",
)
with h5py.File(h5_path, "r") as f:
    var_keys = f["var_key"][:]
    cls = f["cls_ref"][0]   # (10001, 3) softmax track for first variant
    ssu = f["reg_ref"][0]   # (10001,) SSU track for first variant

Download

pip install huggingface_hub
hf download --repo-type dataset mrunyan1/splaire-haec-sqtl-credible-sets